fidopt Interactive view

The datasets

Every dataset of the corpus, by what it is. Choose a task, a kind of image, a region: a card opens the dataset, its results, its source, its licence and its images.

Walk through the images
Task
All415Registration144Segmentation211Synthesis60
Modality
All415CT168MR170CBCT24Ultrasound25PET and SPECT18X-ray5micro-CT4Microscopy35Photographs17
Region
All415Head and neck117Thorax102Abdomen100Pelvis61Spine17Limbs8Cells15Other57
Annotations
All415By people352By a program63
Licence
All415Open257Restricted158
Images
All4152D613D354

Every dataset 415

7725 test cases in 415 datasets. Each picture is the middle slice of the first test case, with the annotations results are checked against in colour; from a dataset's images, Next walks on through every dataset.

What results are checked against. by experts drawn or placed by people. by a program, verified by people found by a program, then checked. by a program the output of a program: a weaker check, and unfair to a method that relies on the same program. a known deformation two images aligned by construction, one of them deformed by a known smooth transform: exact, but on a deformation no body made. as released by a challenge, origin not documented.

Looked at for registration: 1616 public datasets, each with why it is in or out

Every public dataset found that could check a registration, 1616 so far. A dataset enters the main leaderboards when people drew, placed or verified its annotations, when it can be downloaded without a private agreement, and when no other dataset here already holds its images. Each is sorted, with its reason: to build (priority 1 for those that fill a gap of the corpus), waits for a tool (2D, 2D to 3D, video, meshes, images too large to read whole), truth to make, needs an account or a request, or left out.

Thorax · 43

DatasetProblemPairsChecked againstAccessLicenceStatusNotesTo make the truth
4D-Lunglung 4D CT and CBCT, same patient20 patientsone oncologist contoured all 4D CT phases and a few CBCTsboth imagesdirectCC BY 3.0openleft outLearn2Reg ThoraxCBCT, in the corpus, already holds these images. status change: the collection page says 7 of 20 patients have gold coils implanted as fiducials in or near the tumour, visible on 4D CT and 4D CBCT; this gives point truth not noted before. Learn2Reg ThoraxCBCT holds these images. status change: Round 4 count: listed in 2 of the 38 round-4 surveys and benchmark papers (OncoReg 2025; Lung250M-4B 2023). Used in 1 of the 2D and microscopy registration papers read (PolyPose 2025 (as Learn2Reg ThoraxCBCT)).threshold the coils on each phase of 4D CT and 4D CBCT, centroid per coil (or coil end points) = landmark, a person checks the match, 7 patients; limits: coils can migrate, motion blur on CBCT, few points per patient
ACRIN-NSCLC-FDG-PET (ACRIN 6668)NSCLC: FDG PET/CT and CT before and after chemoradiation242 patients; 236 with >=2 studiesCentral SUV reads only (no masks); BAMF AI labels in IDCone imagedirectCC BY 3.0opentruth to make
no truth on both images; PET and CT of one PET/CT scan are aligned by the scanner, so a known warp on the CT gives exact PET-CT truth
Fails truth.PET and CT of one PET/CT scan are aligned by the scanner, so a known warp on the CT gives exact PET-CT truth; for baseline vs post-treatment CT, two people place 20 landmarks (airway and vessel branches, vertebrae) on 10 pairs, TotalSegmentator labels on both as a weaker check.
Anti-PD-1_Lung / Anti-PD-1_MELANOMAPET/CT before and during immunotherapyLung: 46 patients, 40 with 2 studies; Melanoma: 47 patientsNoneone imagedirectCC BY 3.0 (Lung); Melanoma NIH Controlled Data Access Policyopentruth to make
no truth on both images; PET and CT of one PET/CT scan are aligned by the scanner, so a known warp on the CT gives exact PET-CT truth
Fails truth.PET and CT of one PET/CT scan are aligned by the scanner, so a known warp on the CT gives exact PET-CT truth; across visits, two people place 20 skeletal and vessel landmarks on 10 Lung pairs (Melanoma is controlled access), TotalSegmentator labels on both as a weaker check.
Coltea-Lung-CT-100WTriphasic chest CT (native, arterial, venous), same session100 patients x 3 phasesNoneone imagedirectCC BY-SA 4.0 (code and data set)opentruth to make
no truth on both images; Two people place 20 airway bifurcations and bone points on the native, arterial and venous phases of 10 patients
Fails truth criterion Licence checked on 2026-10-10 at https://github.com/ristea/cycle-transformer; derived test pairs shareable: yes; commercial use: yes. README 'License' section; share-alike applies to derived pairs if they are distributedTwo people place 20 airway bifurcations and bone points on the native, arterial and venous phases of 10 patients; phases are close in time, so this tests contrast change more than deformation. Check the licence first.
COPDGene full cohortinspiratory and expiratory breath-hold CT, same subject, plus 5-year follow-upabout 10,000 subjects (unverified)none beyond the 10 DIR-Lab pairs (300 landmarks each, covered)controlledapplication to COPDGene (unverified)askneeds a request
the data needs a request or a data use agreement
Unverified; source of DIR-Lab COPD. Large pool for self-made truth (e.g. Lung250M-style keypoints). status change: usage counted from papers: the training pool of the UNC foundation models; test always on the 10 DIR-Lab COPD pairs. status change: Round 4 count: listed in 6 of the 38 round-4 surveys and benchmark papers (Boveiri 2020 review; Bharati 2022 review; Zou 2022 review (Front Oncol); Chen X 2021 review (Prog Biomed Eng); uniGradICON 2024; multiGradICON 2024).Lung250M-4B recipe on new inspiration/expiration pairs: corrField keypoints with a person verifying a sample on 10 pairs (favours corrField-like methods), plus 50 bifurcations by two people; lungmask lobes on both as a weaker check.
COVID-CTPRED (CREATIS)chest CT of COVID-19 patients: non-contrast, contrast and inspiration/expiration at admission, plus follow-up CTtarget 800 patients; same-day insp/exp and contrast pairs plus follow-ups (counts unverified)nonehiddennot stated; algorithms run on CREATIS computing, images stay thereaskleft out
its images or its truth are not released
Rare same-day insp/exp pairs in a COVID cohort, but images never leave the host.two people place 20 airway/vessel landmarks on 10 insp-exp pairs (DIR-Lab style)
CT Images in COVID-19 (serial subset)Unenhanced chest CT, serial scans of COVID-19 patients29 patients, 121 CTsNoneone imagedirectCC BY 4.0opentruth to make
no truth on both images; Two people place 20 airway and vessel bifurcations away from the opacities on 10 serial pairs
Fails truth.Two people place 20 airway and vessel bifurcations away from the opacities on 10 serial pairs; lungmask lobes on both as a weaker check. Opacities change between scans, so they cannot serve as landmarks.
CT-RATEnon-contrast chest CT with reports, some patients with more than one scan25,692 CTs from 21,304 patients (up to ~4,400 repeat scans; same-day vs follow-up unverified)noneaccountCC BY-NC-SA 4.0, gated; no redistributionnon-commercialneeds an account
the download needs an account
Patient and scan IDs are in file names; repeat intervals must be checked.TotalSegmentator lobes and airways on both (weaker); 20 airway bifurcations by two people on 10 pairs
CT-vs-PET-Ventilation-Imaginginhale and exhale breath-hold CT, free-breathing 4DCT and Galligas PET/CT, same subject same session20 patients: one inhale/exhale BHCT pair, 4DCT phases and PET/CT eachnone spatial (lung function tests only)one imagedirectCC BY 4.0opentruth to make
no truth on both images; Two people place 50 airway and vessel bifurcations on the inhale/exhale breath-hold CTs of 10 patients (skip truncated scans)
Fails truth. Two BHCT pairs show little motion, some scans truncated. Candidate for self-made landmarks. status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Lung250M-4B 2023).Two people place 50 airway and vessel bifurcations on the inhale/exhale breath-hold CTs of 10 patients (skip truncated scans); Galligas PET/CT is hybrid, so a known warp on its CT gives exact PET-CT truth.
DIR-Lab 4DCT and COPDgenelung 4D CT and inspiration to expiration CT, same patient10 + 10300 expert landmarks a pairboth imagesrequestnon-commercial, password after a formnon-commercialleft out
non-commercial terms with a password per requester and no redistribution, so the shared test images the corpus needs are impossible; two requests (2026-10) went unanswered, the lab having left Emory; Lung250M, LungDIRQA, POPI and L2R_NLST cover lung landmark registration
The historical reference. To add when the request is granted. Lung250M-4B already carries the COPDgene landmarks. status change: usage counted from papers: still the reference for lung landmark error, used by learning, implicit-neural and Gaussian methods alike. status change: Round 4 count: listed in 10 of the 38 round-4 surveys and benchmark papers (Fu 2020 review (PMB); Boveiri 2020 review; Bharati 2022 review; Zou 2022 review (Front Oncol); Chen X 2021 review (Prog Biomed Eng); Chen 2025 survey (MedIA); Learn2Reg 2020-21 (TMI 2022); uniGradICON 2024; multiGradICON 2024; Lung250M-4B 2023).
DoseRAD2026 (0.35T MR-Linac MRI with registered CT)0.35 T MR-Linac bSSFP MRI and planning CT, thorax and abdomen, same patient115 patients: 75 released April 2026, 40 test held until 2030none for registration: the CT is already deformably registered to the MRI by the organisers, with manual quality controlone imagedirectCC BY-NC 4.0non-commercialtruth to make
no truth on both images; the pairs are aligned by construction: apply a known smooth warp to the CT for exact MR-CT truth
data hosted on Hugging Face (link on the Zenodo record); a dose challenge, not a registration set; whether the unregistered CT is also shipped is unverifiedthe pairs are aligned by construction: apply a known smooth warp to the CT for exact MR-CT truth; two people place 15 landmarks on 10 pairs to check the given alignment
Duke Lung Cancer Screening Dataset 2024 (DLCS)low-dose screening chest CT with nodule boxesnone: one (terminal) scan per patient; 1,613 public CTssemi-automatic nodule 3D boxesrequestCC BY-NC-ND 4.0 (restricted access, request)non-commercialneeds a request
the data needs a request or a data use agreement
Fails pairs (checked: paper says terminal exam only). Licence checked on 2026-10-10 at https://zenodo.org/api/records/13799069; derived test pairs shareable: unclear; commercial use: no. Zenodo record restricted; ND forbids sharing adapted material (resampled pairs) outside the teamnot applicable (single time point); only usable for synthetic known-warp pairs
EMPIRE10lung CT, same patient30landmarks, fissures, kept by the organisersboth imageshiddenchallenge use onlyaskleft outScoring runs on their server only. status change: Round 4 count: listed in 8 of the 38 round-4 surveys and benchmark papers (Boveiri 2020 review; Bharati 2022 review; Zou 2022 review (Front Oncol); Chen X 2021 review (Prog Biomed Eng); Chen 2025 survey (MedIA); Li 2023 dataset collection (ACM CSUR); Learn2Reg 2020-21 (TMI 2022); Lung250M-4B 2023).
Free-breathing lung MRI raw data (wave-CAIPI 4D; 3D UTE)raw k-space free-breathing lung MRI with 8 breathing phases (wave-CAIPI) or UTE radial, single subjectsa few subjects (unverified)noneone imagedirectCC BY 4.0opentruth to make
no truth on both images; Reconstruct the 8 phases, then two people place 15 vessel bifurcations on end-inhale and end-exhale of each subject
Fails truth; raw data needs reconstruction. Found while checking 4D-MRI lung sets. Licence checked on 2026-10-10 at https://zenodo.org/api/records/3765167; derived test pairs shareable: yes; commercial use: yes. Zenodo record 3765167 (wave-CAIPI 4D); 3D UTE part not checked separatelyReconstruct the 8 phases, then two people place 15 vessel bifurcations on end-inhale and end-exhale of each subject; too few subjects for a test set, a demo case only.
FreeTune4D Digital4D phantomsdigital 4D-MRI phantoms with anatomy labels for motion reconstructionDigital4D-900 benchmark, Digital4D-1M trainingknown motion and labels from the phantomsboth imagesdirectnone stated: abstract only; GitHub repos peilin-wang-git/FreeTune4D and Digital4D900 have no licence and no READMEunknownleft out
no data download location found (the GitHub repos have no README or licence); digital phantoms with simulated motion.
Abstract says datasets are available; download location not found. Licence checked on 2026-10-10 at https://github.com/peilin-wang-git/Digital4D900; derived test pairs shareable: unclear; commercial use: unclear. data download location not found; JBHI paper not open
INSPECT (Stanford pulmonary embolism CT)CT pulmonary angiograms with EHR, same patient at several dates23,248 CTPA from 19,402 patients (project site); first AIMI release 20,078 CTs (~3,800 repeats, unverified)noneaccountresearch data use agreement, non-commercialnon-commercialneeds an account
the download needs an account
Contrast-filled vessels help landmarking.pulmonary vessel bifurcations found by a program then verified by a person on 10 pairs
Learn2Breath (Learn2Reg 2026 task 2)Inspiration/expiration chest CT, same subject (exp moving, insp fixed)200 train + 10 val NLST pairs (1.5 mm, 224x192x224); 100 hidden COPDGene test pairs (>=2 L volume change)Automatic lobe and fissure segmentations on both scans (organisers); test: lobes, fissures, costophrenic surfacesboth imagesdirectno licence; Terms: participants must keep use consistent with the original dataset's data-use terms (named as COPDGene); released train/val files are NLST (TCIA NLST is CC BY 4.0)askin the corpuspriority 2
Its released images are L2R_NLST's; only the lobe/fissure labels are new (add them to L2R_NLST after checking checksums). CodaBench timed out from here: same files inferred from the 200+10 count and grid.
status change for NLST: re-released with lobe+fissure labels on both images. Dataset page says NLST, Terms page says COPDGene-derived (contradiction). Test data never released, Docker only Licence checked on 2026-10-10 at https://www.codabench.org/competitions/15955/; derived test pairs shareable: unclear; commercial use: unclear. Terms (COPDGene, controlled access) contradict the Dataset page (NLST); hidden COPDGene test set never released
LOLA11single chest CT, lung and lobe segmentationnonelung and lobe reference segmentations, hidden (submission only)one imageaccountcustom rules (Zenodo license.txt, 'other-open'): data only for preparing a LOLA11 entry; no other scientific use, no training of other algorithms; no sharing outside the registered teamaskneeds an account
the download needs an account
Fails: no pairs, hidden truth. Licence checked on 2026-10-10 at https://zenodo.org/api/records/4708800/files/license.txt/content; derived test pairs shareable: no; commercial use: no. scans now open on Zenodo 4708800 but the rules forbid using them for other purposes, incl. commercial productsNo second image and hidden labels: no usable registration truth beyond a known smooth warp of a single CT (synthetic).
Longitudinal indeterminate lung nodules (NLST-derived)low-dose CT, same patient, yearly T0 to T2, from NLST443 subjects, up to 3 scansnodule locations per scan from NLST recordsboth imagescontrolledCC BY 4.0 (CSV only)askneeds a request
the data needs a request or a data use agreement
Only a CSV of positions; images must come from NLST under agreement.nodule centres as landmarks plus 20 vessel bifurcations per pair
Lung-Fused-CT-Pathologypre-operative lung CT to whole-slide histology of excised ground-glass nodules, same subject6 patientspathologist-marked invasive and in situ regions on histology, nodule segmentation on CT, creators' registration resultsboth imagesdirectCC BY 3.0openwaits for a tool
whole-slide histology (2D) to CT, and the CT-side correspondence is the creators' own registration; 6 patients.
Tiny; correspondence on CT is the creators' registration. WSI via Aspera.
Lung-PET-CT-DxCT and FDG-PET/CT of lung cancer patients (same-session PET/CT, some separate diagnostic CT)355 subjects, 436 studiestumour bounding boxes by one of five thoracic radiologists, reviewed by all fiveone imagedirectCC BY 4.0opentruth to make
no truth on both images; warp CT by a known field and register PET to it: exact away from lung bases
Free breathing during PET and CT: lung tumours move between modalities. Boxes are on CT slices only.warp CT by a known field and register PET to it: exact away from lung bases; for separate CT vs PET/CT-CT, two people place 20 airway/vessel landmarks on 10 pairs
MIDRAS common datasets (Brock 2010 multi-institution study)lung 4D CT; liver 4D CT and exhale MR to CT; prostate repeat MR; same patientone patient per task (4 tasks)bifurcations and anatomic points localised by radiation oncologistsboth imageshiddennot releasedaskleft out
its images or its truth are not released
The reference multi-institution DIR comparison; its results fed the TG-132 tolerances. Data never public.ask the Deformable Registration Accuracy Consortium (Brock); else vessel bifurcations by two people on public liver 4D CT/MR
MIDRC Data Commonschest radiographs and CT (now also other modalities), multi-site US, many patients with longitudinal records>135,000 public imaging studies; longitudinal subset size unverifiedsome annotations and labels (per study, unverified)accounttwo MIDRC DUAs (University of Chicago): non-commercial research use, or a commercial-use DUA; both forbid distribution, publication or reproduction and derivative works (research results allowed); each user must sign separatelyaskneeds an account
the download needs an account
Part kept sequestered for validation. status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Foundation models review 2026). Licence checked on 2026-10-10 at https://www.midrc.org/midrc-data-use-agreement; derived test pairs shareable: no; commercial use: yes. registration on data.midrc.org; resampled pairs are derivative data, not shareableCheXmask-style lung masks for CXR pairs (weak); 20 landmarks by two people on 10 CT pairs
Multi-breath 4DCT of lung cancer patients (PSI)synthetic multi-breath 4DCT made by transferring motion across 4D-Lung longitudinal scans6 patients, 5 consecutive breaths eachmotion fields used to generate the images (whether shared is unverified)directCC BY 4.0opentruth to make
six synthetic multi-breath 4DCTs, one archive per patient; the record neither lists nor mentions the generating motion fields
Derived from 4D-Lung (listed). Breath-to-breath variability is generated, not measured.if the generating fields are included: dense known truth between breaths; else DIR-Lab style 300 landmarks on 2 patients by two people
NLST AI labels in IDC (TotalSegmentator-CT-Segmentations, BAMF-AIMI, nnU-Net-BPR)Low-dose chest CT of NLST (T0/T1/T2 screens of the same participant)TotalSegmentator labels for most NLST CTs (26,194 participants)Automatic organ segmentations (TotalSegmentator v1.5.6) on every scan; BAMF adds AI lung/nodule labels (~10% radiologist-checked)both imagesZenodo manifests; files in IDC's public AWS and GCS buckets (idc-index or s5cmd, no login)CC BY 4.0 (Zenodo); NLST CT images CC BY 4.0 (TCIA)openleft out
TotalSegmentator v1.5.6 segmentations (DICOM SEG) and radiomics for most NLST CTs, in IDC's buckets through manifests (Zenodo record); BAMF's nnU-Net labels had about 10% of predictions graded by a radiologist, scan by scan (Zenodo 13244892). Labels from a network are what the corpus makes itself (automatic, ranked apart), and NLST pairs are already in the corpus (LUNA25, L2R_NLST, Lung250M).
status change: NLST (in corpus) now has AI organ labels on all timepoints via IDC; BAMF also covers QIN-BREAST, RIDER Lung PET-CT, NSCLC Radiogenomics, ACRIN 6668, Anti-PD-1 Lung, etc. AI labels, not expert truth.
NLST-New-lesion-LongCTLow-dose screening chest CT, same participant: baseline vs follow-up year126 follow-up scans from 122 participants, 152 new lesionsRadiologist point annotation of each new lesion on the follow-up scan; flag whether it was visible on baseline; elastix-registered baseline images providedone imagedirectCC BY 4.0opentruth to make
no truth on both images; Use the original baseline, not the elastix-resampled one
Points on follow-up only (new lesions), so not a two-image landmark set; distinct from the Learn2Reg NLST subset. Published 2025-07.Use the original baseline, not the elastix-resampled one; two people place 20 airway and vessel bifurcations on both scans of 10 pairs, and a reader marks each new lesion flagged visible on baseline at its baseline position.
NLST-Sybil (IDC)Low-dose screening chest CT of NLST participants who developed cancer, several screening years per participant581 participants (pairs per participant not counted)Bounding boxes of suspicious lesions by two thoracic radiologists, including precursor lesions at the cancer location on earlier scansboth imagesZenodo manifests; DICOM boxes and NLST CTs in IDC's public buckets, no loginCC BY 4.0 (Zenodo); NLST CT images CC BY 4.0 (TCIA)openleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): open, but not among the six minor sets kept. It was proposed as: Bounding boxes drawn jointly by two fellowship-trained thoracic radiologists on the LDCTs of NLST participants who developed cancer within a year, precursors at the cancer's location on earlier screens included (Zenodo record, Sybil README); the original JSON keys boxes by image, with no lesion ID across years. A pair would carry one or two box centres of a lesion that grows, under the median of ten landmark pairs; TotalSegmentator labels on both scans (automatic) would cover the lungs, as in LUNA25, which is non-commercial while this is CC BY 4.0.
Sparse lesion correspondences across years (box centres). Hosted by IDC, not TCIA.
NLSTSeg (IDC)Low-dose screening chest CT (NLST)601 participants; number with lesions on >1 year unverifiedExpert volumetric segmentations of suspicious lesions + radiomicsdirectCC BY 4.0opentruth to make
lesions segmented on one scan per participant (the positive screening year), none on the other NLST years; these CTs are also listed under NLST AI labels in IDC
Longitudinal coverage unverified.First count participants with segmented lesions on two years; lesion centroids on both give sparse points, plus 20 bifurcations by two people on 10 pairs and lungmask lobes on both as a weaker check.
NSCLC Radiogenomicschest CT and FDG-PET/CT of NSCLC patients, often both per patient211 subjects, 395 studiestumour segmentations on CT (subset), semantic annotationsone imagedirectCC BY 3.0opentruth to make
no truth on both images; CT to PET/CT-CT: two people place 20 landmarks on 10 pairs
Diagnostic CT and PET/CT are separate sessions: a real CT-CT/PET registration pair. AI organ labels exist via IDC (BAMF).CT to PET/CT-CT: two people place 20 landmarks on 10 pairs; tumour mask warped from CT to PET/CT gives a lesion check; PET-CT within session via known warp
OncoRegCT to CBCT, same patientphase 2 privatelandmarks and labelsboth imageshiddenpublic phase is Learn2Reg ThoraxCBCT; phase 2 data stay in hospitalsaskleft out
its images or its truth are not released
status change: Round 4 count: listed in 2 of the 38 round-4 surveys and benchmark papers (Learn2Reg 2024 overview; OncoReg 2025).
Pseudo-CBCT from AAPM thoracic CT (Zenodo 5035494)planning CT converted to simulated CBCT with geometric augmentation (synthetic CT-CBCT pairs)23 CT/psCBCT/organ-mask pairs per AAPM thoracic CT (patient count unverified)organ masks (oesophagus, heart, lungs, cord) carried through the augmentation, on bothboth imagesdirect: Zenodo, two zips (4.7 and 17.3 GB) whose members read by HTTP ranges (listed and headers read that way)CC BY 4.0 on the Zenodo record (the source LCTSC is CC BY 3.0 on TCIA)openin the corpuspriority 3
The record holds the 60 LCTSC patients (24 test patients in test_psCBCT_AAPM.zip, 36 training patients in train_psCBCT_AAPM.zip), each as the planning CT resampled to 0.98 mm and cropped (159x163x239 for Test-S1-101), a pseudo-CBCT simulated from it on the same grid (OS-SART reconstructions at several settings) and the four LCTSC contours. The training set's geometric augmentations move CT, psCBCT and masks together and are not saved, so no CT-to-CBCT displacement exists: the truth would be a known warp drawn by the corpus (synthetic). The patients are the corpus's LCTSC (its draw pairs all 60, 40 of them in its test and development pairs) and the CBCT is simulated, so it ranks apart; real thorax CBCT/CT with people's points is L2R_ThoraxCBCT.
Synthetic CBCT; whether the geometric transforms are saved is unverified.known warp: keep or re-run the published augmentation to get exact CT-CBCT displacement truth
Pulmonary tree labeling (airway, artery, vein point graphs)pulmonary tree graphs from CT, between patientsbetween shapesbranch labelsboth imagesdirectApache-2.0 (repository)openleft out
binary airway/artery/vein shapes and graphs, no CT images listed; branch labels would have to be mapped back to CTs.
Shape set; labelled branches could serve as point truth for lung CT registration. Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Point-cloud shape survey 2025).map branch labels back to the CTs: labelled bifurcations as landmarks between patients
Real-world multimodal longitudinal lung cancer dataset (MMIST-LUNG)CT, PET and whole-slide images over follow-up, lung cancer1,365 patients, many missing modalitiesnoneone imagerequestno licence on the MMIST repo; CT/PET come from TCIA TCGA-LUAD/LUSC (CC BY 3.0) and CPTAC-LUAD/LSCC (CC BY 3.0/4.0)openneeds a request
the data needs a request or a data use agreement
Access conditions not stated. Licence checked on 2026-10-10 at https://multi-modal-ist.github.io/datasets/Lung/; derived test pairs shareable: yes; commercial use: yes. MMIST site says clinical variables harmonised from CPTAC and TCGA; take images straight from TCIA; one CPTAC-LSCC head CT is NIH controlledlung and lobe labels from a public segmenter on both CTs (weaker)
RIDER Lung PET-CTRepeat PET/CT244 subjectsUnverifiedone imagedirectCC BY 3.0opentruth to make
no truth on both images; PET and CT of one PET/CT scan are aligned by the scanner, so a known warp on the CT gives exact PET-CT truth
Fails truth; BAMF AI labels exist in IDC. Licence checked on 2026-10-10 at https://www.cancerimagingarchive.net/collection/rider-lung-pet-ct/; derived test pairs shareable: yes; commercial use: yes. TCIA collection table; cite the DOI and follow the TCIA Data Usage PolicyPET and CT of one PET/CT scan are aligned by the scanner, so a known warp on the CT gives exact PET-CT truth; for repeat CTs, two people place 20 landmarks on 10 pairs, TotalSegmentator labels on both as a weaker check.
S0819-Tumor-Annotations (on S0819 (SWOG NSCLC))CT and PET/CT of advanced NSCLC, same patient across trial timepoints (baseline, on-treatment, follow-up)1,279 patients annotated; 1,126 with >=2 annotated studies; 3,744 lesions tracked (1,113 patients)NCI annotation project: up to 5 RECIST/PERCIST lesions per scan segmented, same lesions tracked over time (DICOM Tracking ID/UID) plus seed points; radiologists, US board-certified reviewboth imagescontrolledannotations CC BY 4.0; images NCTN/NCORP licence, now NIH Controlled Data Access Policyaskneeds a request
the data needs a request or a data use agreement
Sparse lesion correspondences (centroids/seed points with tracking IDs) and lesion masks on both scans; lesions change size so use as landmarks, not dense overlap. Largest tracked-lesion set found. NCTN trial images: TCIA lists them as "Unavailable" since July 2025 (moving to NCI CTDC, dbGaP request, trial-specific PHS); annotations themselves are public.
SegTHOR (thoracic organs at risk CT)chest CT, between patients40 training scans with labels (60 total)manual heart, aorta, trachea, oesophagus labelsboth imagesaccountSegTHOR Data Use Agreement (accepted on CodaLab registration): research only, no commercial use, no passing to third parties; new purposes need CHB approval; no public postingnon-commercialleft out
access by a signed agreement; the organisers' forum reads it as forbidding redistribution (checked 2026-10-16)
Segmentation set used for inter-patient overlap. Licence checked on 2026-10-10 at https://pagesperso.litislab.fr/cpetitjean/wp-content/uploads/sites/19/2021/11/DataUseAgreement_CHBDatabaseV3.pdf; derived test pairs shareable: unclear; commercial use: no. team members must each sign the DUA; CodaLab terms: 'You are not allowed to redistribute the data'two people place 15 landmarks (carina, aortic arch) on 20 pairs
SPARE challenge (sparse-view 4D-CBCT)1-minute 4D-CBCT projections plus prior 4DCT, same subject; reconstruction with motionMonte Carlo set 12 patients/32 scans; Varian 5 patients/30 scans; Elekta 6 patients/24 scansground-truth 4D-CBCT images and evaluation scripts; no landmarksone imagedirectno formal licence: 'Anyone is welcome to use the SPARE Challenge datasets for their research', cite Shieh et al.non-commercialtruth to make
no truth on both images; In the Monte Carlo set the CBCT is simulated from a known 4DCT, so landmarks placed by two people on that CT and on the prior 4DCT carry exactly to the CBCT
Fails registration truth. Monte Carlo set built from TCIA 4D-Lung (covered). Repository hdl.handle.net/2123/34378 returned 403 to the fetcher. status change: Round 4 count: listed in 2 of the 38 round-4 surveys and benchmark papers (Hammoudeh 2023 survey; Chen X 2021 review (Prog Biomed Eng)). Licence checked on 2026-10-10 at https://image-x.sydney.edu.au/spare-challenge/; derived test pairs shareable: unclear; commercial use: unclear. data on Sydney eScholarship hdl.handle.net/2123/34378 (repository page blocked to scripts, rights field not read); commercial use not addressedIn the Monte Carlo set the CBCT is simulated from a known 4DCT, so landmarks placed by two people on that CT and on the prior 4DCT carry exactly to the CBCT; clinical sets get 15 bifurcations by two people per patient.
SPARK (TROG 15.01)CT and kV projection images–not volume pairsrequestnot readable: listed on Image X Institute 'Public Data' page, record on Sydney eScholarship (hdl 2123/31090) blocked to scriptsunknownneeds a request
the data needs a request or a data use agreement
Licence checked on 2026-10-10 at https://image-x.sydney.edu.au/public-data/; derived test pairs shareable: unclear; commercial use: unclear. repository record returns 403 to curl and WebFetch; check the rights field in a browserImplanted gold seeds show on the planning CT and on every kV projection: seed centroids on CT, projected through the recorded gantry geometry, give 2D/3D marker truth checked against the KIM traces. Licence unverified.
SPIROMICSinspiratory and expiratory CT of COPD subjects, same subject, longitudinalunverifiednone knowncontrolledunverifiedaskneeds a request
the data needs a request or a data use agreement
Unverified; listed as the other big inspiration/expiration COPD cohort.Same recipe as Lung250M-4B: corrField keypoints on inspiration/expiration pairs with a person verifying a sample (favours corrField-like methods), plus 50 bifurcations by two people on 10 pairs; lungmask lobes as a weaker check.
SPREAD (Continuous Registration Challenge private set)Longitudinal chest CT (lung densitometry study), same subjectPrivate set; size not on page (unverified)Landmarks (per Continuous Registration paper, unverified)both imageshiddenPrivateaskleft out
its images or its truth are not released
Named as the only private one of 8 evaluation sets; other 7 already covered status change: Round 4 count: listed in 2 of the 38 round-4 surveys and benchmark papers (Boveiri 2020 review; Zou 2022 review (Front Oncol)).
Synthetic 4DCT Mid-Position benchmarksynthetic 4DCT made from 20 clinical 4DCT, with reference Mid-P images20 synthetic 4DCTknown synthetic motion and 4D tumour masksboth imagesrequestnone found: paper (Phys Med 2023, CC BY article) abstract gives no data release or licenceunknownneeds a request
the data needs a request or a data use agreement
Release not confirmed. Licence checked on 2026-10-10 at https://doi.org/10.1016/j.ejmp.2023.103144; derived test pairs shareable: unclear; commercial use: unclear. full text on ScienceDirect blocked to scripts; release not confirmed
VAMPIRE challenge dataset (CT ventilation)4DCT inhale/exhale DIR to derive ventilation, compared with Galligas PET, DTPA-SPECT or Xenon-CT, same subject50 subjects (25 Galligas, 21 DTPA-SPECT humans, 4 Xenon-CT sheep)reference ventilation images, lung masks, SIFT-detected landmark pairs (automatic)both imagesrequestnone stated in the paper; no public download found (screening note: by request under IRB data-sharing arrangements)askneeds a request
the data needs a request or a data use agreement
Indirect truth (function correlation); landmarks are automatic, not expert. Licence checked on 2026-10-10 at https://pmc.ncbi.nlm.nih.gov/articles/PMC6605778; derived test pairs shareable: unclear; commercial use: unclear. paper (Kipritidis 2019, Med Phys) says an on-line dataset was intended but gives no URL or licence
VESSEL12single chest CT, lung vessel segmentationnone (20 single scans)vessel/non-vessel point labels, hidden except 3 example filesone imagedirectCC BY 4.0opentruth to make
no truth on both images; No second image: no usable registration truth beyond a known smooth warp of a single CT (synthetic).
Fails: no pairs. Listed because it was asked.No second image: no usable registration truth beyond a known smooth warp of a single CT (synthetic).

Cardiac · 28

DatasetProblemPairsChecked againstAccessLicenceStatusNotesTo make the truth
5D cardiac MRI on a 1.5 T MR-linac (CMR-MOTUS k-space set)ungated 3D cardiac MR k-space on an MR-linac, reconstructed cardiorespiratory motion states with deformation fields, 10 volunteers; compared with 2D cine10 subjects x many motion statesphysical and digital phantom validation in the paper; in vivo only consistency with 2D cine motion; reconstructed DVFs are method outputone imagedirectCC BY 4.0opentruth to make
no truth on both images; LV and RV labels on motion states by a public segmenter
Model DVFs are not independent truth; 2D cine acquired tooLV and RV labels on motion states by a public segmenter; two people mark mitral annulus and apex on end states
CAP LV Segmentation 2011cine MR200myocardium labelsboth imagesrequestCAP Data Use Agreement + challenge terms: data only for the cardiac segmentation challengeaskneeds a request
the data needs a request or a data use agreement
Licence checked on 2026-10-10 at https://www.cardiacatlas.org/lv-segmentation-challenge/; derived test pairs shareable: no; commercial use: unclear. FAQ: 'Can I use the data for other project than the cardiac segmentation? No.'; CAP user policy: no transfer of data or derivatives without CAP approval
CETUS3D echo ED to ES, same patient15 to 45contours by 3 expertsboth imagesaccountCC BY-NC-SA 4.0 (zea docs)non-commercialneeds an account
the download needs an account
CMRxMotioncine MR ED to ES and breath-hold to free-breathing, same subject20 training volunteersmanual masks (only on good-quality scans)both imagesaccountCC BY non-commercialnon-commercialneeds an account
the download needs an account
CMRxRecon (2023, 300 subjects)cine MR (SAX stack and 2/3/4-chamber) and T1/T2 mapping, raw multi-coil k-space plus images, healthy volunteers, frames and weightings of the same slice300 subjects (120 train fully sampled)manual chamber and myocardium labels by one radiologist on cine SAX/LAX and mapping (phases labelled unverified)both imagesaccountCC BY ('the CMRxRecon dataset follows the CC-BY license')openneeds an account
the download needs an account
Also a motion-correction problem for T1/T2 mapping weightings with labels; later CMRxRecon2024/2025 add multi-centre data, labels unverified Licence checked on 2026-10-10 at https://repo-prod.prod.sagebase.org/repo/v1/entity/syn52965326/wiki; derived test pairs shareable: yes; commercial use: yes. Synapse wiki; free Synapse account, no application; CC-BY version not givenlabels on mapping weightings give label truth; RV insertion points by two people on 30 subjects
CMRxRecon 2023, 2024, 2025 (cardiac multi-contrast k-space)cardiac cine, T1/T2(/T2*) mapping, tagging, phase contrast, black blood, LGE, perfusion (2025); same subject same session~300 (2023) + 330 (2024) healthy volunteers, single centre; ~600 (2025) multi-centre with patientsnone for registrationaccountfree for research and education; commercial use of the dataset prohibited (2024 paper)non-commercialneeds an account
the download needs an account
Each sequence is a separate breath-hold, so slices of different sequences do not line up exactly. Within one mapping series the TI/TE-weighted images share slice and breath-hold.within a T1 or T2 mapping series, warp one weighted image by a known 2D field and register another to it: exact multi-contrast truth; tagging vs cine of the same slice for a weaker check
Continuous-time 4D cardiac CT built by coronary-mask-guided registrationmultiphase cardiac CT turned into artifact-reduced continuous-time 4D sequencespatient count and release unverifiedpseudo ground truth made by registration (circular for registration tests)one imagehiddenunverifiedunknownleft out
its images or its truth are not released
Release of the dataset not confirmed.use as known-motion phantom only: the warps are known, so they give exact truth for re-registration
In-house 4D cardiac CT cohorts (AMC TAVI 1,509; MGH 61 4D CCTA; Ruijin 18)ECG-gated 4D CCTA, 10-21 phases per patient, same patient over the cardiac cycleAMC 1,509 patients (21,605 volumes); MGH 61 sequences x 20 phases; Ruijin 18 x 10AMC: 3D chamber labels at ES/ED for 12 patients, 2D slices in 4 phases for 81; MGH: LV/myocardium on 247 frames; Ruijin: LV cavity on all 10 phases by technician and clinicianboth imageshiddenin-houseaskleft out
its images or its truth are not released
public 4D cardiac CT remains almost absent (only the pig set and TIB listing found earlier)if shared: chamber labels in every phase give label truth; valve annulus and coronary ostia points by two people
Kaggle Second Annual Data Science Bowl (cardiac MRI)2D cine MR short-axis stacks plus 2/4-chamber, 1,140 patients, frames of the same slice1,140 studies, ~30 frames per sliceonly ED and ES LV volumes per patient; no contours releasedone imageaccountKaggle competition rules (accept to download)non-commercialneeds an account
the download needs an account
Fails truth; huge and varied; volumes check only gross label overlappublic cardiac segmenter (nnU-Net on ACDC/M&Ms) on all frames, a weaker label check; two people click RV insertions on 30 subjects
LVQuan18 (DIG-Cardiac, processed)2D cine MR, single mid-ventricular short-axis slice, 20 frames over a cycle, patients from 3 Canadian hospitals145 training subjects x 20 frames (test 30 subjects without truth)endo- and epicardial myocardium masks on every frame (one engineer, checked by two cardiac radiologists) plus wall thickness, dimensionsboth imagesrequestnone stated (challenge site, data via Baidu/Dropbox links)unknownneeds a request
the data needs a request or a data use agreement
Dense per-frame myocardium labels on a single 2D slice: good label truth for 2D cine registration; data via challenge contact (unverified now) Licence checked on 2026-10-10 at https://lvquan18.github.io/; derived test pairs shareable: unclear; commercial use: unclear. no licence or terms on the pagetwo people place 6 myocardial points (RV insertions, AHA segment borders) on ED and ES of 30 subjects
LVQuan192D cine MR, mid short-axis slice, 20 frames, original unprocessed images56 training subjects x 20 frames (1,120 images); 30 test without truthmyocardium masks on every frame plus LV indices (engineer, checked by two radiologists)both imagesdirectnone stated (training data by e-mail 'to count the number of participants', test data on Google Drive/Baidu)unknownwaits for a tool
2D single-slice cine MR; training data sent by e-mail.
Training data on Google Drive/Baidu per GitHub mirror (TiancongHua repo); same source cohort as LVQuan18 Licence checked on 2026-10-10 at https://lvquan19.github.io/; derived test pairs shareable: unclear; commercial use: unclear. no licence or terms on the pageRV insertion points clicked by two people on all 20 frames of 20 subjects
M&Ms and M&Ms-2cine MR ED to ES, same patient375 and 360manual ED and ES labelsboth imagesaccountM&Ms Dataset Research Use Agreement: research use only, sale prohibited, non-academic research needs case-by-case approval; no distribution or derivative works without written permissionaskleft out
M&Ms Research Use Agreement (read 2026-10-16), clause 5: 'You may not distribute, publish, or reproduce a copy of any portion or all of the M&Ms Dataset to others without specific prior written permission'; a written permission from the organisers (Lekadir, Campello) would reopen it
status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Chen 2025 survey (MedIA)). Licence checked on 2026-10-10 at https://www.ub.edu/mnms/documents/data_use_agreement.pdf; derived test pairs shareable: no; commercial use: no. same agreement for M&Ms-2; Canadian centre data (30 cases) only via collaboration
MITEA3D echo ED to ES and scan-rescan, same subject134 subjects, 536 imageslabels derived from cardiac MR (semi-automatic)both imagesrequestCC BY-NC-SA 4.0, no redistributionnon-commercialneeds a request
the data needs a request or a data use agreement
MM-WHSCT and MR, inter-subject20 + 20 training7 structures, manualboth imagesrequestregistration form agreement: data disseminated for the MM-WHS 2017 analysis; recipients commit to not disseminate to any third partyaskleft out
MM-WHS terms (read 2026-10-16): 'The Recipient(s) commit to not disseminate the Data to any third party'; maintenance ended
status change: Round 4 count: listed in 6 of the 38 round-4 surveys and benchmark papers (Bharati 2022 review; Hammoudeh 2023 survey; Chen X 2021 review (Prog Biomed Eng); Chen 2025 survey (MedIA); Azad 2023 ViT review; Li 2023 dataset collection (ACM CSUR)). Licence checked on 2026-10-10 at https://zmiclab.github.io/zxh/0/mmwhs/; derived test pairs shareable: unclear; commercial use: unclear. signed form sent to organisers; purpose tied to the challenge
MS-CMRSeg 2019bSSFP, LGE, T2 MR, heart, same patient45manual LV, RV, myocardiumboth imagesrequestsigned Data Use Agreement Document; recipients commit to not disseminate the data to any third partyaskneeds a request
the data needs a request or a data use agreement
Licence checked on 2026-10-10 at https://zmiclab.github.io/zxh/0/mscmrseg19/; derived test pairs shareable: unclear; commercial use: unclear. data link sent after signed agreement
Multi-view echocardiography sequences (Peressutti)3D echo of LV from different acoustic windows, same volunteer4 volunteers x up to 5 windowsRigid transforms from optical tracking of the probeboth imagesdirect: Zenodo record 30999, one tar.gz a sequence (not readable by ranges, 30 to 80 MB each)CC0 1.0 (Zenodo)openleft outpriority 2
The truth is optical probe tracking, not people, and the files contradict themselves: a pair's i_to_j and j_to_i transforms are not inverses under the authors' own convention (gaps up to 35 mm), and loops over three windows miss by 3 to 37 mm.
Multimodal STRAUSsimulated 3D US, cine MR, tagged MR of the same virtual patient18 patients, 90 sequencesreference meshes from an electromechanical model (synthetic), every frameboth imagesdirectnone stated (public CREATIS collection, no licence file)unknownleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): its licence is not shown to allow commercial use (none stated (public CREATIS collection, no licence file)). It was proposed as: Synthetic 3D echo, cine MR and tagged MR of 18 virtual patients with model meshes every frame: fills the cardiac US gap with simulated truth (mesh point correspondence not stated); no licence stated; 13 GB.
Licence checked on 2026-10-10 at https://humanheart-project.creatis.insa-lyon.fr/database/api/v1/collection/587de6f4e1af3f30a2980a58; derived test pairs shareable: unclear; commercial use: unclear. collection description and multimodalityStraus.html carry no terms
Multiphase cardiac CT scans dataset (TIB listing)ECG-gated multiphase cardiac CT, same subject across cardiac phases50 mitral valve prolapse + 50 coronary disease patients (unverified)unverifiedrequestnot readable: TIB LDM page and CKAN API behind a bot check (DOI 10.57702/rz053ven)unknownneeds a request
the data needs a request or a data use agreement
Unverified: the page is behind a bot check. DOI 10.57702/rz053ven. Only 4D cardiac CT lead found. Licence checked on 2026-10-10 at https://service.tib.eu/ldmservice/dataset/multiphase-cardiac-ct-scans-dataset; derived test pairs shareable: unclear; commercial use: unclear. TIB LDM indexes datasets named in papers; the source host is not identifiedTwo people place 15 landmarks (coronary ostia and bifurcations, valve hinge points, apex) on end-diastole and end-systole of 10 patients; open TotalSegmentator heart labels on both phases as a weaker check.
MVAA 2026 (mitral valve CT, 3D TEE and surgical video)cardiac CT, intraoperative 3D transesophageal echo and surgical video of the mitral valvethree separate segmentation tasks; whether the same patients appear in all modalities is unverifiedexpert mitral valve and leaflet segmentations per modalitydirectunverifiedunknownwaits for a tool
the proposal pairs only 40 CTs with surgical video, with no shared truth (video to 3D: waits for a 2D/3D pose error); it names no CT-TEE pairs; data CC BY-NC
Training data on Google Drive (Springer community post). Likely fails pairs if cohorts differ.if any patients have both CT and TEE, annulus landmarks (commissures, saddle points) placed by two people on both
MyoPS 2020bSSFP, LGE, T2 cardiac MRI, same patient45 patients, 2-6 slices eachLV/RV/myocardium/scar/edema labelsboth imagesaccountMyoPS Data Access Agreement: data for the sole purpose of the challenge analysis, cannot be used in another context; no third-party dissemination; delete within 36 months after Sept 2020askneeds an account
the download needs an account
Fails: sequences pre-aligned (MvMM) before release Licence checked on 2026-10-10 at https://zmiclab.github.io/zxh/0/myops20/files/DataAccessAgreementforParticipants_MyoPS.pdf; derived test pairs shareable: no; commercial use: no. signed by recipients and a permanent faculty member
OCMR (open-access multi-coil k-space cardiac cine)2D cardiac cine MR raw k-space (fully sampled and prospectively undersampled), frames of the same slice165 fully sampled + 212 undersampled series (279 + 842 slices)none (reconstruction dataset)one imagedirectOCMR terms and conditions: research and educational purposes only; derived datasets bound by the same terms; no selling or monetisingnon-commercialtruth to make
no truth on both images; public cardiac segmenter labels on all frames
Fails truth; includes 0.55T data; useful to test registration on reconstructions with motion artefacts Licence checked on 2026-10-10 at https://www.ocmr.info/download/; derived test pairs shareable: unclear; commercial use: no. click-through terms before the download pagepublic cardiac segmenter labels on all frames; two people click RV insertions on ED/ES
PointCHD (congenital heart disease point clouds)heart point clouds from CT, between patientsbetween shapespart segmentationsboth imagesrequestdataset by e-mail from the author; README says source ImageCHD 'following the CC-BY license'askneeds a request
the data needs a request or a data use agreement
Shape set (JBHI 2024). Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Point-cloud shape survey 2025). Licence checked on 2026-10-10 at https://github.com/Dinghow/PointCHD; derived test pairs shareable: unclear; commercial use: unclear. no LICENSE file, no data licence; contact dinghowyang@gmail for download
PULSE (pulmonary artery deformation finite-element meshes)right ventricular outflow tract and pulmonary artery surfaces from cardiac MR, reference and deformed shapes, same patient60 patients, 900 deformed configurationsfinite-element displacement with point-to-point correspondenceboth imagesdirectunverifiedunknownwaits for a tool
finite-element meshes and point clouds only, no images.
Sci Data 2026. Meshes and point clouds only, no images; data host not opened.render the meshes into synthetic volumes to make image pairs with exact displacement
STACOM 2014 perfusion MRI motion correction challengefirst-pass perfusion MR time series (one short-axis slice), rest and stress, same patient9-10 cases x 2 series x ~50 framesmanual myocardial contours by a single observer; perfusion indicesboth imagesaccountrequest form + CAP Policies for Data Distribution to Users (use solely for the approved Research Project; no transfer of data or derivatives; delete at project end)askneeds an account
the download needs an account
Listed in Li 2023. Small but a real 2D+time registration benchmark with evaluation code (Pontre 2017). Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Li 2023 dataset collection (ACM CSUR)). Licence checked on 2026-10-10 at https://www.cardiacatlas.org/wp-content/uploads/2022/10/CAPPolicyStatementUsers.pdf; derived test pairs shareable: no; commercial use: unclear. challenge page links a request form; no licence nameda second observer redraws contours on 3 cases to measure observer error
STONE free-breathing myocardial T1 mapping (BIDMC)2D short-axis T1-mapping series (11 frames), same subject, groupwise motion correction210 subjects, 1,050 slicesexpert LV and myocardium annotations; RV insertion points used for TREboth imagesdirectno STONE download page found; the related MyoMapNet in-vivo T1 maps (MOLLI/LL4, BIDMC) on Harvard Dataverse are CC0 1.0unknownwaits for a tool
2D T1-mapping series, and no download page found.
Described as public in Any2Reg (El-Rewaidy, Nezafat et al.); download page not found. A related public MOLLI set is on Harvard Dataverse (MyoMapNet, doi:10.7910/DVN/5MZYAH). Licence checked on 2026-10-10 at https://dataverse.harvard.edu/api/datasets/:persistentId/?persistentId=doi:10.7910/DVN/5MZYAH; derived test pairs shareable: unclear; commercial use: unclear. CC0 applies to the Dataverse set doi:10.7910/DVN/5MZYAH, which is not the STONE free-breathing data; STONE itself has no terms found
STRAUS 3D synthetic echocardiography (Alessandrini 2015)simulated 3D echocardiography sequences (B-mode and RF) from an electromechanical heart model, healthy and ischemic, frame-to-frame8 sequences, one cycle eachground-truth myocardial meshes over the cycle (exact simulated motion)both imagesdirect: Apache directory listing at http://www-sop.inria.fr/asclepios/data/STRAUS/ (Bmode/, meshes/, RF/), no accountnone stated; the page says 'the use of such data is citation-ware' (cite Alessandrini et al., IEEE TMI 2015)askin the corpuspriority 2
The open directory www-sop.inria.fr/asclepios/data/STRAUS/ serves the 8 simulated 3D echocardiography sequences (normal, ladprox, laddist, lcx, rca, lbbb, lbbbsmall, sync: healthy, ischaemic and dyssynchronous) as B-mode volumes (94 to 125 MB zipped each), RF signals and the ground-truth myocardial meshes of each frame. The truth is the simulated motion (exact) but only 8 sequences exist, and the page states no licence, only that use 'is citation-ware' (cite Alessandrini et al., TMI 2015): ranked apart.
3D, not 2D; distinct from Multimodal STRAUS (Zhou 2018) already listed; download at www-sop.inria.fr/asclepios/data/STRAUS/
Synthetic healthy and infarcted cine MR from STACOM 2011 (Duchateau 2018)synthetic cine MR sequences generated from 15 real STACOM 2011 cases with model-driven motion (1 healthy + 30 infarct variants each), frame-to-frame15 x 31 synthetic sequences plus the real inputssimulated meshes driving the image warps (exact known motion)both imagesdirect: Apache directory listing at http://www-sop.inria.fr/asclepios/data/STRAUS/SyntheticMRI/, no accountnone stated; 'COPYRIGHT NOTICE: published reports ... should cite' Duchateau et al. 2018 and Tobon-Gomez et al. 2013askin the corpuspriority 3
The open directory www-sop.inria.fr/asclepios/data/STRAUS/SyntheticMRI/ holds 15 volunteers of STACOM 2011 (v1 to v16), each with the original cine short-axis sequence (InputImages) and 1 healthy plus 30 infarcted synthetic sequences made by warping it with model-driven motion (SyntheticData, VTK images and meshes). The truth is the simulated motion, 15 subjects cannot give ten test and ten development, and the page only asks to cite Duchateau et al. (TMI 2018) and Tobon-Gomez et al. (MedIA 2013). The volunteers are those of the corpus's STACOM2011 (tagged MR there, cine here), and cardiac cine is covered with people-made truth by ACDC, Sunnybrook and YorkCardiacMR.
Real image texture warped by known motion: exact cine registration truth; derived from STACOM 2011 (listed) but a distinct release
Tagged MRI at 0.55T example dataset2D cardiac tagged MR at 0.55T, frames of the same sliceexample subjects only (count unverified)none stated; tag grid intersections are physical markers in the tissueone imagedirectCC BY 4.0opentruth to make
no truth on both images; tag-line intersections tracked by two people (as Loecher 2021 did, 11-50 points per slice) on a few slices
One of very few public tagged-MR sets besides STACOM 2011; tags fade over the cycletag-line intersections tracked by two people (as Loecher 2021 did, 11-50 points per slice) on a few slices

Breast · 23

DatasetProblemPairsChecked againstAccessLicenceStatusNotesTo make the truth
ACRIN-FLT-Breast (ACRIN 6688)FLT PET/CT before and during neoadjuvant chemotherapy83 patients; 74 with >=2 studiesNoneone imagedirectCC BY 3.0opentruth to make
no truth on both images; PET and CT of one PET/CT scan are aligned by the scanner, so a known warp on the CT gives exact PET-CT truth
Fails truth.PET and CT of one PET/CT scan are aligned by the scanner, so a known warp on the CT gives exact PET-CT truth; across visits, two people place 15 skeletal and vessel landmarks on 10 pairs, TotalSegmentator labels on both as a weaker check.
Auckland prone and supine breast MRI of volunteers (ABI breast biomechanics)T1-weighted 1 mm MRI of the same women prone and supine110 volunteers (plus 86 prone-only patients)soft-tissue landmarks tracked between poses by the group (theses); skin, rib cage, nipple segmentationsboth imageshiddenin-house, not releasedaskleft out
its images or its truth are not released
no public prone-supine breast MRI found anywhere; this is the largest known cohortif shared: fibroglandular branch points and cysts clicked by two radiologists on 10 subjects
BREAST-CANCER-SCREENING-DBT (TCIA, re-checked)DBT craniocaudal and mediolateral oblique views of the same breastCC-MLO view pairs; lesion boxes for 201 patientslesion bounding boxes by radiologistsdirectCC BY-NC 4.0non-commercialleft outpriority 3
One box centre a pair: 153 of the 180 breasts with boxes on both views have one box a view (boxes CSVs of the training, validation and test sets), and when there are several the boxes are not linked across views; too thin to rank methods. CC BY-NC.
No prior exams documented, so still no public temporal DBT pairs; only cross-view boxes.CC and MLO are different projections of a compressed breast, so no dense truth; lesion box centres on both views give one matched point per lesion, a weak check only.
Dartmouth Prone to Supine Breast MRI trials (NCT03573804, NCT05131607)diagnostic prone and supine breast MRI of the same patients for surgical planningabout 62 patients planned in the first trialsurgical and pathology correlation; skin markers (unverified)hiddentrial data, not releasedaskleft out
its images or its truth are not released
also UPM Madrid prone-to-supine surface workflow on 18 cases (Alfano 2019), in-houseskin fiducials placed before both scans give exact surface points; lesion centres by a radiologist
DMID (Digital mammography Dataset for Breast Cancer Diagnosis Research)mammograms with reports and lesion masksno pairs foundlesion masksone imagedirectCC BY 4.0opentruth to make
no truth on both images; Single exams: only between patients
MGRegBench: lacks paired images. Fails pairs: one image per subject, no repeat or second modality. Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (MGRegBench 2025).Single exams: only between patients; a known smooth deformation on one mammogram, or pair CC and MLO views with lesion centres as weak points.
EA1141Dense-breast screening: abbreviated breast MRI and digital breast tomosynthesis, yearly rounds500 women; 498 with >=2 studies (360 with 4)None seenone imagedirectCC BY 4.0opentruth to make
no truth on both images; Biopsy clips, where present, are markers on later MRI rounds
Fails truth; MR/DBT multimodal and longitudinal pairs. Published 2026-05.Biopsy clips, where present, are markers on later MRI rounds; otherwise two people place 10 landmarks (nipple, vessel branches) on 10 year-to-year MRI pairs. MRI vs DBT compression differs too much for dense truth.
HDMI breast photoacoustic + ultrasound CT (Dryad)photoacoustic CT and ultrasound CT of the breast from one dual-modal system, same scan161 breast tumours (322 MAT files)pathology diagnosis onlydirectCC0opentruth to make
322 MAT files, photoacoustic and ultrasound CT of 161 tumours from one ring array (co-registered by construction), pathology only; the page does not say whether a file is a slice stack or one slice
Both modalities come from one ring system at 10 Hz: aligned by construction (verify registration claim in the paper).warp the ultrasound CT by a known field and register photoacoustic to it: exact multimodal truth (vessels vs tissue)
I-SPY1, ACRIN-6698, MAMA-MIA, Duke breast MRIbreast MR before and during chemotherapyhundredstumour masks only; the tumour shrinks or a single time pointone imagedirectCC BY 3.0 / 4.0opentruth to make
no truth on both images; Biopsy clips placed before chemotherapy show as signal voids on later MRIs: clip centroid per visit, checked by a person, gives marker truth
Biopsy clips placed before chemotherapy show as signal voids on later MRIs: clip centroid per visit, checked by a person, gives marker truth; two people add 10 vessel branches and the nipple on 10 pairs.
ISPY2 (+ BreastDCEDL_ISPY2)Longitudinal breast DCE-MRI during neoadjuvant therapy (unverified)719 patientsUnverifiedboth imagesdirectCC BY 4.0 (images and supporting files)opentruth to make
the row gives no truth on both images; 1.75 TB, beyond the laptop.
I-SPY1 covered; this is the larger follow-up trial Licence checked on 2026-10-10 at https://www.cancerimagingarchive.net/collection/ispy2/; derived test pairs shareable: yes; commercial use: yes. TCIA collection table; BreastDCEDL_ISPY2 derivative not checked separately
KAU-BCMD (King Abdulaziz University mammograms)mammograms, same breast and view across screening exams for some women1,416 cases, 2,206 JPG images; pairs in MGRegBenchnone for registration (BI-RADS, density)directCC0: Public Domain (Kaggle, uploaded by the author asmaasaad)opentruth to make
no truth on both images; MGRegBench already adds expert landmarks to some pairs
Source of MGRegBench pairs. Repository found by search; official status unverified. Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (MGRegBench 2025). Licence checked on 2026-10-10 at https://www.kaggle.com/datasets/asmaasaad/king-abdulaziz-university-mammogram-dataset; derived test pairs shareable: yes; commercial use: yes. Kaggle API licenseName; GitHub repo Asmaaalsolami/KAU-BCMD is GPL-3.0 with an empty README; paper (Data 2021) article CC BY 4.0MGRegBench already adds expert landmarks to some pairs; more: two readers place 10 landmarks on 20 pairs
LAMIS-DMDB (digital mammography database)mammogramsunverifiedunverifiedrequestsigned LAMIS-DMDB Dataset Transfer Agreement (Univ. Tébessa): research only, no commercial work; confidential, no disclosure or distribution to third parties without written consent; return or dispose after the researchnon-commercialneeds a request
the data needs a request or a data use agreement
Named by MGRegBench as not publicly accessible. Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (MGRegBench 2025). Licence checked on 2026-10-10 at https://lab-lamis-univ-tebessa.dz/docs/agreemennt_LAMISDBM%20(1).docx; derived test pairs shareable: unclear; commercial use: no. agreement template found via a Law Insider listing; signed by the recipient organisationCheck for prior exams first; if any, two radiologists mark lesion centres and calcifications on 10 prior/current pairs.
Mammo-Bench19,731 mammograms from 6 public sets, ~6,500 patientsonly where a source set has prior examslesion masks, BI-RADSone imageaccountCC BY-NC-SA 4.0non-commercialneeds an account
the download needs an account
Aggregate of public sets; MGRegBench had to inspect each source for pairs. Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (MGRegBench 2025).Find patients with prior exams across the 6 sources; two radiologists mark masses and calcifications on 10 pairs.
MammoTrackermammograms over timeunverifiedlesion trackingboth imagescontrolledneeds EMBED accessaskneeds a request
the data needs a request or a data use agreement
MGRegBenchmammography prior to current, same view100 landmarked + 5,907 unlabelled~10 landmarks a pair by 2 radiologists, breast masksboth imagesinherits INbreast, KAU-BCMD and RSNA termsunknownleft outThe repository given in the paper returns 404 (2026-10-10). status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (MGRegBench 2025).
PAM3 hybrid photoacoustic + ultrasound breast tomography (Twente)3D multi-wavelength photoacoustic and speed-of-sound reconstructions from one scanner, phantom and in vivofew phantoms and volunteers (unverified)reference SOS model for a test objectdirectMIT (as on Zenodo)opentruth to make
3D photoacoustic and speed-of-sound volumes (MAT) of 8 volunteers and test objects from one scanner, aligned by construction; no registration truth, a known warp would make one
Reconstructions only (figure data); PA and SOS share geometry by construction.warp the SOS volume by a known field and register a PA wavelength to it: exact truth; different wavelengths give extra contrasts
PriorNet DBT dataset (Duke)DBT current exam to 1-year prior, same subject973 cancer and 7,123 non-cancer cases from 8 siteslesion annotations on current exams and prior locationsboth imageshiddenunknownleft out
its images or its truth are not released
No public access found; ask the authors.
QIN Breast DCE-MRIBreast DCE-MRI at visits during neoadjuvant chemotherapy10 patients, 2 visits eachNone (pathological response only)one imagedirectCC BY 3.0opentruth to make
no truth on both images; Two people place 10 landmarks (nipple, vessel branches, clips) on both visits of all 10 patients
Fails truth.Two people place 10 landmarks (nipple, vessel branches, clips) on both visits of all 10 patients; pre- and post-contrast frames of one visit under a known warp give near-exact multi-phase truth.
QIN-BREASTBreast cancer: prone FDG PET/CT (built for MRI registration) and quantitative prone MRI at 3 timepoints of neoadjuvant therapy68 patients, 216 studiesNone from submitters (BAMF AI labels in IDC)one imagedirectCC BY 3.0opentruth to make
no truth on both images; PET and CT of one PET/CT scan are aligned by the scanner, so a known warp on the CT gives exact PET-CT truth
Fails truth; good PET/CT-MRI prone multimodal pairs.PET and CT of one PET/CT scan are aligned by the scanner, so a known warp on the CT gives exact PET-CT truth; for prone PET/CT to prone MRI, two people place 10 landmarks (nipple, vessel branches, sternum) on 10 pairs.
QIN-BREAST-02breast MR at several visits–no annotationsone imagecontrolledrestrictedaskneeds a request
the data needs a request or a data use agreement
Restricted access; if obtained, two people place 10 landmarks (nipple, vessel branches, biopsy clips) on 10 visit pairs.
RIDER Breast MRIBreast DWI MRI, coffee-break repeat and interval exams5 patients, 2 studies eachNoneone imagedirectCC BY 3.0opentruth to make
no truth on both images; Coffee-break repeats have near zero motion: identity and inverse consistency checks only
Fails truth; tiny.Coffee-break repeats have near zero motion: identity and inverse consistency checks only; 5 patients is too few for anything else.
RSNA Screening Mammography Breast Cancer Detectionscreening mammograms, ~20,000 women54,705 images; same-breast pairs where prior exams existnone for registrationaccountKaggle competition rules; not redistributable (MGRegBench)non-commercialneeds an account
the download needs an account
Source of MGRegBench pairs (preprocessing script only). Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (MGRegBench 2025).two readers place landmarks on 20 pairs
Sequential mammograms with breast masses (Loizidou, Cyprus)prior and recent full-field mammograms, CC and MLO, same patient100 pairsevery mass delineated by two expert radiologists on both imagesboth imagesdirectCC BY 4.0opentruth to make
prior and recent mammograms (CC BY, Zenodo, 3.3 GB) whose truth is the masses drawn by two radiologists, one structure a breast; two readers place 10 landmarks on 20 pairs to reach a main board (mass masks kept as other labels)
MGRegBench names it as a candidate extension; sister set of the micro-calcification pairs already listed. Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (MGRegBench 2025).mass masks match across time; add 10 landmarks per pair by two readers for a point check
Sequential micro-calcification mammogramsfull-field mammograms prior to recent, same view, same patient200expert mask of every calcification on both images, no correspondence IDsboth imagesdirectCC BY 4.0opentruth to make
same Cypriot source and design; calcification masks carry no identities: match clusters by program, a person checks, or add reader landmarks; build after the masses set
weak truth status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (MGRegBench 2025).

Abdomen · 55

DatasetProblemPairsChecked againstAccessLicenceStatusNotesTo make the truth
20 MRI-CT pairs for colorectal liver metastasis ablationliver MRI to CT, same patient20 pairsunverified (no description)requestCC BY 4.0 (restricted access)openneeds a request
the data needs a request or a data use agreement
Restricted record with no description.vessel and tumour-centre landmarks by two people
3D-IRCADb-02 (respiratory cycle)CT arterial phase in inhale to portal phase in exhale, same patient1organ, vessel and lesion segmentations and meshes on both scans (IRCAD)both imagesdirectCC BY-NC-ND 4.0non-commercialto buildpriority 3
One pair only (inhale arterial vs exhale portal CT); ND licence.
One pair only; ND clause may block derived labels.
4DCT-MRI liver digital phantoms (Ribeiro 2018, PSI/UMCG)liver 4D CT made by warping a 3D CT with motion from 4D MRI, same patient9 data setsground-truth DVFs from 4D MRIboth imageshiddennot releasedaskleft out
its images or its truth are not released
Same 4DCT(MRI) idea as the listed Duetschler lung set; liver sets not found online.request from PSI; the method itself makes exact truth from any 4D MRI plus CT
AbdomenAtlas 1.1CT, inter-patient9,262 CT25 structures labelled by a network then revised by radiologistsboth imagesaccountCC BY-NC-SA 4.0non-commercialneeds an account
the download needs an account
AbdomenCT-1KCT, inter-patient1,0004 organsboth imagesrequestno data licence on repo (Apache-2.0 covers code); download via Google data usage tracking form; FLARE22 Zenodo calls AbdomenCT-1K labels 'research purpose only'askleft out
its images are mostly LiTS, KiTS19, MSD and NIH Pancreas-CT, already in the corpus; its tumour labels are a model's; the new 13-organ part is FLARE22 (Zenodo 7860267)
Licence checked on 2026-10-10 at https://github.com/JunMa11/AbdomenCT-1K; derived test pairs shareable: unclear; commercial use: unclear. images re-collected from MSD, KiTS, NIH Pancreas, LiTS, each with its own terms
AHEP0731-Tumor-Annotations (on AHEP0731 (COG hepatoblastoma))CT, MR of pediatric liver tumours, same patient across trial timepoints (baseline, on-treatment, follow-up)79 patients annotated; 41 with >=2 annotated studies; 74 lesions trackedNCI annotation project: up to 5 RECIST/PERCIST lesions per scan segmented, same lesions tracked over time (DICOM Tracking ID/UID) plus seed points; radiologists, US board-certified reviewboth imagescontrolledannotations CC BY 4.0; images NIH Controlled Data Access Policyaskneeds a request
the data needs a request or a data use agreement
Sparse lesion correspondences (centroids/seed points with tracking IDs) and lesion masks on both scans; lesions change size so use as landmarks, not dense overlap. Images downloadable only with dbGaP approval (TCIA Data Retriever + CRDC key).
AMPLIFAI (multi-phase liver CT, MICCAI 2026)multi-phase contrast CT of the liver (non-contrast, arterial, portal venous, delayed), same patient same session590 public cases (531 train, 59 validation) with 2-4 phases; 174 private test casesone radiologist lesion mask per case plus feature masks on specific phases (APHE on arterial, capsule on venous/delayed)one imageaccountunverified (challenge terms of use)unknownneeds an account
the download needs an account
Curated from WAW-TACE, HCC-TACE-Seg, TCGA-LIHC and PLC-CECT, so it overlaps listed sources but adds re-annotation.two people place 10 vessel bifurcation pairs between arterial and venous phases on 20 cases; liver masks from a public segmenter on each phase as a weaker check
AREN0532-Tumor-Annotations (on AREN0532 (COG))CT, some MR of pediatric renal tumours (Wilms), same patient across trial timepoints (baseline, on-treatment, follow-up)540 patients annotated; 69 with >=2 annotated studies; 75 lesions trackedNCI annotation project: up to 5 RECIST/PERCIST lesions per scan segmented, same lesions tracked over time (DICOM Tracking ID/UID) plus seed points; radiologists, US board-certified reviewboth imagescontrolledannotations CC BY 4.0; images NCTN/NCORP licence, now NIH Controlled Data Access Policyaskneeds a request
the data needs a request or a data use agreement
Sparse lesion correspondences (centroids/seed points with tracking IDs) and lesion masks on both scans; lesions change size so use as landmarks, not dense overlap. NCTN trial images: TCIA lists them as "Unavailable" since July 2025 (moving to NCI CTDC, dbGaP request, trial-specific PHS); annotations themselves are public.
AREN0533-Tumor-Annotations (on AREN0533 (COG))CT of pediatric Wilms tumour (incl. lung metastases), same patient across trial timepoints (baseline, on-treatment, follow-up)293 patients annotated; 238 with >=2 annotated studies; 455 lesions trackedNCI annotation project: up to 5 RECIST/PERCIST lesions per scan segmented, same lesions tracked over time (DICOM Tracking ID/UID) plus seed points; radiologists, US board-certified reviewboth imagescontrolledannotations CC BY 4.0; images NCTN/NCORP licence, now NIH Controlled Data Access Policyaskneeds a request
the data needs a request or a data use agreement
Sparse lesion correspondences (centroids/seed points with tracking IDs) and lesion masks on both scans; lesions change size so use as landmarks, not dense overlap. NCTN trial images: TCIA lists them as "Unavailable" since July 2025 (moving to NCI CTDC, dbGaP request, trial-specific PHS); annotations themselves are public.
AREN0534-Tumor-Annotations (on AREN0534 (COG))CT, MR of pediatric bilateral/familial Wilms tumour, same patient across trial timepoints (baseline, on-treatment, follow-up)236 patients annotated; 227 with >=2 annotated studies; 513 lesions trackedNCI annotation project: up to 5 RECIST/PERCIST lesions per scan segmented, same lesions tracked over time (DICOM Tracking ID/UID) plus seed points; radiologists, US board-certified reviewboth imagescontrolledannotations CC BY 4.0; images NCTN/NCORP licence, now NIH Controlled Data Access Policyaskneeds a request
the data needs a request or a data use agreement
Sparse lesion correspondences (centroids/seed points with tracking IDs) and lesion masks on both scans; lesions change size so use as landmarks, not dense overlap. NCTN trial images: TCIA lists them as "Unavailable" since July 2025 (moving to NCI CTDC, dbGaP request, trial-specific PHS); annotations themselves are public.
B-mode-and-CEUS-Liver (TCIA)B-mode and contrast-enhanced ultrasound cine of liver lesions, same subject120 subjectsnone spatial; pathology and cross-sectional imaging diagnosis onlyone imagedirectCC BY 4.0opentruth to make
no truth on both images; If the cine is dual-display (B-mode and CEUS side by side in one plane), the two are aligned by construction and a known warp gives exact multimodal truth
Fails truth. Only useful for CEUS motion compensation if someone adds landmarks.If the cine is dual-display (B-mode and CEUS side by side in one plane), the two are aligned by construction and a known warp gives exact multimodal truth; for frame-to-frame motion, two people track 5 vessel points on 10 clips.
BFH liver CT (Beijing Friendship Hospital, RCN)abdominal CT, between patients92 scansnonedirectnone stated for the data (repo MIT covers code)unknowntruth to make
no truth on both images; TotalSegmentator liver labels on both, weaker check
Fails truth; training pool only. Licence checked on 2026-10-10 at https://github.com/microsoft/Recursive-Cascaded-Networks; derived test pairs shareable: unclear; commercial use: unclear. README only asks to cite; preprocessed MSD+BFH on Google DriveTotalSegmentator liver labels on both, weaker check
BTCV Multi-Atlas Labeling Beyond the Cranial Vault: Abdomenportal-venous contrast CT of the abdomen, between subjects (label overlap)30 labelled training scans (870 directed pairs); 20 test scans with labels held by the organisers13 organs labelled by two trained students and verified by a radiologist (MIPAV)both imagesaccountno named licence: 'access under the conditions of the data sponsors'; register on Synapse and agree to the licence termsaskleft out
its images are in the corpus through Learn2Reg AbdomenCTCT
Source of the corpus's L2R AbdomenCTCT. Xu 2016 used 100 scans from the same two Vanderbilt trials; only 50 are released. Test labels stay with the organisers; scoring service status unverified. status change: source of Learn2Reg AbdomenCTCT (in the corpus) and pre-training data of its MR-CT task. Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Learn2Reg 2020-21 (TMI 2022)). Licence checked on 2026-10-10 at https://repo-prod.prod.sagebase.org/repo/v1/entity/syn3193805/wiki/217753; derived test pairs shareable: unclear; commercial use: unclear. Synapse Data wiki and 2013 Data Description ('register and agree to the license terms... Do not send the direct link to anyone')labels are expert: add vessel landmarks by two people on 20 pairs
CancerVerselongitudinal CT, up to 26 studies, 13 abdominal cancersthousands (unverified)tumour masks reviewed by radiologists; matching across times not documentedboth imagesdirectCC BY-NC-ND 4.0non-commercialleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): its licence is not shown to allow commercial use (CC BY-NC-ND 4.0). It was proposed as: 3 TB, beyond the laptop; the tumour masks change over time and their matching across visits is not documented; ND licence.
ND: derived labels cannot be shared
CARE-Liver 2026 (LiQA multi-sequence liver MRI)liver MRI with T2, DWI, T1 and Gd-EOB-DTPA arterial, venous, delayed and hepatobiliary phases, same patient, fibrosis800 patients (190 more than CARE 2025)liver mask on the hepatobiliary phase only; other sequences unlabelled (organisers suggest registration-based segmentation)one imageaccountunverifiedunknownneeds an account
the download needs an account
status change: LiQA (CARE 2024, listed) grew to 800 patients and the 2026 segmentation task now scores non-contrast sequences, which rewards cross-sequence registration. Registration closed.two people trace the liver on T2 and T1 for 20 patients and place 10 vessel landmarks per patient across sequences
CHAOS CTabdomen CT and MR–manual liver and organ labelsdirectCC BY-NC-SA 4.0non-commercialleft outCT and MR come from different patients; see CHAOS MR T1 and T2. status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Han 2022 GAN registration review).CT and MR come from different patients: only inter-subject CT-to-MR, scored by the manual liver labels on both images (contour check).
CPTAC-CCRCC-Tumor-Annotations (on CPTAC-CCRCC)CT, MR of clear-cell renal carcinoma, same patient across trial timepoints (baseline, on-treatment, follow-up)59 patients annotated; 7 with >=2 annotated studies; 14 lesions trackedNCI annotation project: up to 5 RECIST/PERCIST lesions per scan segmented, same lesions tracked over time (DICOM Tracking ID/UID) plus seed points; radiologists, US board-certified reviewboth imagesdirect: TCIA NBIA API (open), annotations and source images series by seriesCC BY 4.0 (annotations and source images, TCIA analysis-result page)openin the corpuspriority 3
In the corpus as segmentation; only 7 patients have two annotated studies, too few for a registration set.
Sparse lesion correspondences (centroids/seed points with tracking IDs) and lesion masks on both scans; lesions change size so use as landmarks, not dense overlap. Few longitudinal cases.
CPTAC-PDA-Tumor-Annotations (on CPTAC-PDA)CT, MR of pancreatic cancer, same patient across trial timepoints (baseline, on-treatment, follow-up)99 patients annotated; 8 with >=2 annotated studies; 10 lesions trackedNCI annotation project: up to 5 RECIST/PERCIST lesions per scan segmented, same lesions tracked over time (DICOM Tracking ID/UID) plus seed points; radiologists, US board-certified reviewboth imagesdirect: TCIA's NBIA API or Data Retriever manifest, no accountCC BY 4.0 (annotations and the CPTAC-PDA images, TCIA pages)openin the corpuspriority 3
In the corpus as segmentation; 8 longitudinal patients only.
Sparse lesion correspondences (centroids/seed points with tracking IDs) and lesion masks on both scans; lesions change size so use as landmarks, not dense overlap. Few longitudinal cases.
CT-ORGCT, inter-patient140organ labelsboth imagesdirect: IBM Aspera public link on the collection page, no account (as CTORG's prepare.py fetches it)CC BY 3.0 (TCIA page); the 20 3D-IRCADb-01 scans are CC BY-NC-ND 4.0 on IRCAD's page and stay outopenin the corpuspriority 3
The corpus already holds CT-ORG as CTORG (segmentation, ten of the 21 test scans): a registration set reuses that download, but its moving images are resampled after the affine alignment, so it is built, not derived. The 21 test scans (140 patients, one scan each) have every organ drawn or corrected by hand and reviewed, enough for 10 test pairs of 20 patients, the training scans' lungs and bones (algorithm only) serving development; inter-patient abdominal CT is already covered by AMOS_CTCT, L2R_AbdomenCTCT, PancreasCT8 and PediatricCT.
CTpred-Sunitinib-panNETPancreatic NET: contrast CT before sunitinib and at first follow-up38 patients x 2 CTsNone (response class only)one imagedirectCC BY 4.0opentruth to make
no truth on both images; Two people place 15 landmarks (vessel branches, vertebrae, kidney poles) on both CTs of 10 patients
Fails truth.Two people place 15 landmarks (vessel branches, vertebrae, kidney poles) on both CTs of 10 patients; TotalSegmentator labels on both as a weaker check.
Digital-twin gastrointestinal motionSynthetic 4D GI motion from static CT/MRI, same patient10 patients x 20 phasesKnown DVFs, deformed masks (MSKCC pipeline)both imagesdirect: Zenodo record 17702490, one zip a patient, readable by rangesCC BY 4.0 (Zenodo)openleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): open, but not among the six minor sets kept. It was proposed as: Zenodo (CC BY 4.0, 38.6 GB, one 2.2 to 5.5 GB zip a patient): 11 patients (six T2w MRI, two T1w 4D MRI, three contrast-enhanced CT, PMB 2025), each a static scan and 20 synthetic phases of analytical GI motion with their DVFs and deformed masks, float64 401x401x250 images at 1 mm. The truth is the provider's synthetic deformation, whose generator is published, and 11 patients cannot give ten test and ten development subjects; abdominal same-patient registration already has people-made landmarks (AbdomenDIRQA, LiverDIRQA).
Duke Liver MRImulti-sequence liver MR, same patient310 series, 95 patientsmanual liver masksrequestCC BY-NC-ND 4.0non-commercialneeds a request
the data needs a request or a data use agreement
Sequences from one acquisition (Dixon in/opposed phase) are aligned by construction: a known warp gives exact multi-contrast truth; across breath-holds, liver masks where drawn on both plus MRSegmentator labels as a weaker check.
ETH 4D liver MRI4D liver MR, same subject14 subjectsvessel landmarks for 7 subjectsboth imagesrequestno public release or terms found (ETH BMIC/CVL dataset pages return 404)unknownneeds a request
the data needs a request or a data use agreement
Licence checked on 2026-10-10 at https://bmic.ee.ethz.ch/research/datasets.html; derived test pairs shareable: unclear; commercial use: unclear. von Siebenthal-type 4D liver MRI; likely by request from ETH authors
FLARE22CT, inter-patient50 labelled CT13 organsboth imagesdirectCC BY 4.0 (Zenodo); organ annotations from AbdomenCT-1K 'research purpose only'askin the corpuspriority 2
In the corpus as a segmentation dataset (FLARE22); its CTs are decathlon pancreas scans, and inter-patient abdominal CT registration is already AMOS_CTCT's, PancreasCT8's and L2R_AbdomenCTCT's.
Licence checked on 2026-10-10 at https://zenodo.org/api/records/7860267; derived test pairs shareable: yes; commercial use: unclear. Zenodo record says CC BY 4.0 but the description restricts AbdomenCT-1K labels to research; images from MSD (CC BY-SA 4.0)
Free-breathing liver MRI with motion fields (PTB, MRpro)3D golden radial phase encoding liver MRI raw data, four respiratory states, one subject1 subject, 4 motion statesdisplacement fields between states (estimated by the authors)directCC BY 4.0opentruth to make
one subject's raw k-space (ISMRMRD, to reconstruct with MRpro) in four breathing states; the shipped displacement fields are the authors' estimates, not truth
Fields are estimates, not truth; k-space needs reconstruction. Earlier version Zenodo 15288250.One subject: two people place 15 vessel bifurcations on each motion state to check the authors' fields; a demo case, not a test set.
H108M liver multiphase CT (Vietnam)multiphase liver CT, same patient136 evaluation pairsunverifiedhiddenprivate (partner hospital)unknownleft out
its images or its truth are not released
Used by TotalRegistrator 2025 (Pham et al. 2024 data). Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (TotalRegistrator 2025).phases of one exam: liver and vessel masks by a program checked by a person
HCC-TACE-Segmulti-phase CT, liver cancer, same patient105labels on the pre-procedure scan onlyone imagedirectCC BY 4.0opentruth to make
no truth on both images; Two people place 15 vessel-branch and bone landmarks on pre- and post-procedure CT of 10 patients
Two people place 15 vessel-branch and bone landmarks on pre- and post-procedure CT of 10 patients; TotalSegmentator liver and vessel labels on both as a weaker check.
ImFusion public ultrasound demo: liver sweep to MRItracked 2D B-mode liver sweep and MRI of one healthy volunteer, ultrasound-to-MRI1 volunteernone (demo of ImFusion registration)one imagedirectCC BY 4.0 (per repo)opentruth to make
no truth on both images; two people mark 10 vessel bifurcations in the sweep and the MRI
Single case demo; other folders of the repo may hold more US datatwo people mark 10 vessel bifurcations in the sweep and the MRI
KiTS23CT, kidney, inter-patient489kidney, tumour, cystboth imagesdirectCC BY-NC-SA 4.0non-commercialin the corpuspriority 2
In the corpus, ranked apart: 11 test and 10 development pairs of the 226 patients neither KiTSCT nor KiTS23 holds, scored on TotalSegmentator's organs (automatic), the provider's kidneys shown.
status change: Round 4 count: listed in 2 of the 38 round-4 surveys and benchmark papers (Zou 2022 review (Front Oncol); Hammoudeh 2023 survey).
LiQA (CARE 2024 liver fibrosis track)multi-phase multi-sequence liver MRI (T2, DWI, arterial, portal, delayed, hepatobiliary), same subject, phases not registered, multi-centre440 patients, 6-7 sequences eachliver masks on hepatobiliary phase only (10 manual per centre, rest nnU-Net pseudo-labels); fibrosis stageone imagedirectCC BY 4.0 (per dataset paper)opentruth to make
no truth on both images; Two people place 10 landmarks (portal and hepatic vein branches, liver dome, kidney poles) on the hepatobiliary and portal phases of 10 patients per centre
Real inter-phase misalignment by design; would need labels on a second phase.Two people place 10 landmarks (portal and hepatic vein branches, liver dome, kidney poles) on the hepatobiliary and portal phases of 10 patients per centre; TotalSegmentator MRI liver labels on every phase as a weaker check.
LiTS (Liver Tumor Segmentation Challenge)contrast CT of the liver, between patients131 training scans (131 x 130 pairs in RCN)manual liver and tumour masksboth imagesaccountCC BY-NC-SA 4.0 (unverified)non-commercialneeds an account
the download needs an account
Inter-patient liver overlap only; 131 cases overlap with MSD Task03 (liver).two people place 6-10 liver landmarks (portal vein bifurcation, hepatic vein confluence) on 20 pairs, as RCN did on SLIVER07
LLD-MMRImulti-phase liver MR, same patient498 patientslesion labels (per phase unconfirmed)requestCC BY-NC-ND 4.0non-commercialneeds a request
the data needs a request or a data use agreement
If lesion boxes exist on each phase, lesion centres give one point per phase; two people add 10 vessel-branch landmarks across phases on 10 patients, MRSegmentator liver labels as a weaker check.
Longitudinal abdominal CT with expert-validated liver and metastasis masks (Zenodo 2026)contrast CT, liver metastases under chemotherapy, same patient over time54 patients, several scans each (dates and intervals in metadata.csv)liver and metastasis masks on every scan, expert-validatedboth imagesdirectCC BY 4.0openleft out
Already in the corpus as LiverMetsCT: the same 54-patient release (Dataverse TUW8FZ); this Zenodo copy is restricted.
found through the Zenodo API; lesions shrink or grow under therapy, so lesion masks are not a fixed reference; total size not listedtwo people place 20 liver vessel bifurcations on 10 scan pairs, lesions excluded; liver mask overlap as a second check
Longitudinal pre- and post-treatment liver CT for image registration (Isfahan)portal-venous CT before and after chemotherapy, same patient64 patients, 128 CTexpert-validated liver masks at both time pointsboth imagesrequestCC BY 4.0 (restricted access)openneeds a request
the data needs a request or a data use agreement
Restricted record; overlaps the Harvard Dataverse set.vessel bifurcation landmarks found by a program then verified
MCT-LTDiag (four-phase liver CT, 5 tumour types)non-contrast, arterial, portal venous and delayed CT of the liver, same session517 patientsexpert tumour masks (phase coverage unverified)directunverified (Harvard Dataverse)unknowntruth to make
tumour and liver masks on the portal venous phase only (case archive checked), the other phases registered to it with elastix B-spline per the paper; CC0 on Dataverse
Sci Data 2025.two people place 10 vessel bifurcation pairs across phases on 20 cases
Merlin abdominal CT + Merlin Plus longitudinal metadataabdomen-pelvis CT with reports (Stanford 2012-2018), same patient at several dates25,494 CTs from 18,317 patients; 3,830 patients with 2+ scans (Merlin Plus)AI masks of 44 organs, vessels and sub-segments on all scans; 1,193 radiologist tumour masksboth imagesaccountCT: Stanford AIMI non-commercial research terms; Merlin Plus files: 'other' (unverified)non-commercialneeds an account
the download needs an account
Largest longitudinal body CT pool found with labels on both sides; AI labels are weak truth.two people place 20 landmarks (vessel branchings, vertebrae) on 10-20 pairs; AI organ labels give Dice on all 3,830 patients
Morfeus validation cohort (Velec 2017, Princess Margaret)thoracic and abdominal 4D CT and MR, liver CT-MR, prostate MR; same patient74 patientsinternal point landmarks (TRE) and structure surfaces (DTA)both imageshiddennot releasedaskleft out
its images or its truth are not released
One of the largest landmark-validated clinical DIR cohorts; not public.request from UHN Toronto
OpenMultiModalLiverCirrhosisDatasetliver MRI, CT and ultrasound volumes stated as co-registered, per case1,080 casesnone for registration (fibrosis stage labels only)directCC BY 4.0opentruth to make
no truth on both images; none credible until provenance is confirmed
no institution, ethics or acquisition details on the record; 1,080 trimodal cases with 3D ultrasound in 1.3 GB looks synthetic; verify provenance before any usenone credible until provenance is confirmed
OVGU 4D liver MRI4D liver MR, same subjectseveralnonedirectCC BY 4.0opentruth to make
no truth on both images; Two people place 15 vessel bifurcations on end-inhale and end-exhale of each subject (inter-observer error reported)
Two people place 15 vessel bifurcations on end-inhale and end-exhale of each subject (inter-observer error reported); few subjects, extra test cases only.
PANTHERMR-linac T2 and diagnostic T1, pancreasunknown whether fractions are pairedmanual pancreas and tumourrequestCC BY-NC 4.0non-commercialneeds a request
the data needs a request or a data use agreement
Check that MR-linac and diagnostic scans pair up; where they do, the manual pancreas and tumour on both are contours on both images, and two people add 10 landmarks (vessel origins, kidney hila) on 10 pairs.
PanTrackportal-venous CT series, 2 to 11 per patient, pancreas cancer and liver metastases116 matched lesion pairs, 45 patientsmanual lesion masks with tracking IDs (one radiologist)both imagesdirectCC BY-NC 4.0non-commercialin the corpuspriority 2
PanTrack: the first linked pair of 20 patients (10 test, 10 development), ranked on the provider's TotalSegmentator organs, the lesion centres (median 1 a pair) shown; PanTrackCT takes the other 20 patients.
Princess Margaret liver SBRT breathing-motion dose accumulation cohort (Velec 2011)free-breathing liver SBRT: static exhale planning CT and breathing-phase images (details unverified), same patient21 patientsnone beyond clinical contours (dose accumulation study); MORFEUS registrationhiddennot releasedaskleft out
its images or its truth are not released
Not public.request; vessel bifurcations on exhale and inhale by two people
Princess Margaret MR inhale/exhale volunteers (Brock 2005 MORFEUS)thoracic and abdominal MR of healthy volunteers at inhale and exhale, same subjecta few volunteers (count unverified)visible bifurcations tracked in lung and liverboth imageshiddennot releasedaskleft out
its images or its truth are not released
Not public.request; vessel bifurcations by two people
Siemens/Mayo tracked ultrasound to CT, liver and kidney lesions (Wein 2008)tracked freehand 3D ultrasound to contrast CT, liver and kidney, same patient25 patientsphysician-defined ground truth registrations and target lesionsboth imageshiddennot releasedaskleft out
its images or its truth are not released
Classic CT-ultrasound study; not public.request; vessel bifurcations on US and CT by two people
Simulated post-operative liver resection registration benchmark (Duke Liver MRI)3D liver MRI with a simulated resection cavity, same patient pre vs simulated post10 patientsknown simulated resection and deformation (synthetic)both imagesdirectCC BY 4.0openleft out
only result tables and code deposited; the synthetic resections must be rebuilt on the Duke Liver Dataset.
only derived results and code are deposited; images must be rebuilt from the Duke Liver Dataset under its own termsrebuild the simulation with the released code so the applied warp is known exactly
SLIVER07contrast CT of the liver, between patients20 training scans with masks (10 test, masks kept by the organisers)manual liver masksboth imagesaccountsite rules: data only for developing liver segmentation algorithms; other use needs the maintainers' explicit permission; no distribution outside the registered team; no commercial use of algorithms trained on itaskneeds an account
the download needs an account
Inter-patient liver only; mirrors on Zenodo and Hugging Face. status change: RCN (Zhao et al. ICCV 2019) reports 6 anatomical keypoints on every SLIVER07 training scan, placed by 3 expert doctors and averaged; preprocessed 128^3 crops with these landmarks are shared via the RCN repository's Google Drive link (licence of the derived files not stated). status change: Round 4 count: listed in 4 of the 38 round-4 surveys and benchmark papers (Bharati 2022 review; Hammoudeh 2023 survey; Chen X 2021 review (Prog Biomed Eng); Han 2022 GAN registration review). Licence checked on 2026-10-10 at https://sliver07.grand-challenge.org/Rules/; derived test pairs shareable: no; commercial use: no. registration benchmarking is 'other purposes' and needs permissionuse the RCN 6-keypoint set after checking it; add 10 more vessel landmarks by two people on the 20 scans
SNMMI Lu-177 dosimetry challenge (multi-time-point SPECT/CT)Lu-177 DOTATATE SPECT/CT at 4 time points over ~1 week, plus planar scans, baseline diagnostic scans, TIA maps2 patients x 4 SPECT/CT time pointsorgan (technologist) and lesion (radiologist) contours at each time pointboth imagesdirectCC BY-NC 4.0 (Zenodo mirror, e.g. records/17271433)non-commercialleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): its licence is not shown to allow commercial use (CC BY-NC 4.0 (Zenodo mirror, e.g. records/17271433)). It was proposed as: SPECT/CT time points are new to the corpus, but only 2 patients (4 time points each).
Each SPECT has its own co-registered CT (hybrid, aligned up to breathing); no registration was done between time points in the released data. Very small.CT-CT between time points: two people place 20 landmarks; organ contours give a Dice check; within one time point warp CT and register SPECT for exact truth
SynthRAD2025CBCT and planning CT, MR (incl. 0.35 T MR-Linac) and CT, same patient; head and neck, thorax, abdomen RT patients, 5 centres2362 cases: 1472 CBCT-CT, 890 MR-CT; training (about 65%) publicnone; rigid alignment by organisers; deformed CT for validation and test held until 2030one imagedirectCC BY-NC 4.0 (centre D data: challenge-only licence, to be deleted after)non-commercialtruth to make
no truth on both images; screen CT for metal markers (threshold, then a person confirms), centroid on CT and CBCT/MR = landmark
Paper says metal implants are present in a limited number of cases. Liver/pancreas fiducials, surgical clips and lung coils are likely in the abdomen and thorax subsets but unverified. 1x1x3 mm resampling.screen CT for metal markers (threshold, then a person confirms), centroid on CT and CBCT/MR = landmark; count marker cases first; limits: 3 mm slices, CBCT streaks bias centroids, markers can migrate
TCGA radiology collections with repeat scans (TCGA-KIRC, -LIHC, -LUAD)clinical CT and MR of cancer patients, same patient at several datesKIRC: 46 of 237 CT patients with 2+ CTs; LIHC: 38 of 75 CT and 16 of 40 MR; LUAD: 34 of 67 CT (NBIA API)none on pairs; BAMF AIMI automatic segmentations linked for TCGA-KIRC (others unverified)directCC BY 3.0 (TCGA-KIRC page; other collections unverified)opentruth to make
no truth on both images; TotalSegmentator or BAMF organ labels on both
Heterogeneous clinical protocols (phase, field of view). status change: Round 4 count: listed in 2 of the 38 round-4 surveys and benchmark papers (Foundation models review 2026; Learn2Reg 2020-21 (TMI 2022)).TotalSegmentator or BAMF organ labels on both; two people place 20 landmarks on 10 pairs per organ
TriALSmulti-phase liver CTunverifiedunverifiedaccountCC BY-NC 4.0 (HF dataset marwankefah/TriALS, gated with manual approval); repo code Apache-2.0non-commercialneeds an account
the download needs an account
Licence checked on 2026-10-10 at https://huggingface.co/api/datasets/marwankefah/TriALS; derived test pairs shareable: yes; commercial use: no. arXiv 2605.16572 points to github.com/xmed-lab/TriALS for data; uses MSD Liver images tooCheck that the phases are not pre-registered; two people place 15 vessel-branch landmarks across phases on 10 patients, TotalSegmentator liver and vessel labels on all phases as a weaker check.
TRUSTED3D transabdominal ultrasound to CT of the kidney, same patient48 patients, 96 kidneys7 landmarks per kidney and kidney masks by 2 radiographers (STAPLE)both imagesrequestData Usage Agreement (data_use_agreement.txt) to sign and e-mail before download; paper says researchers may use it freely to develop and validate methodsaskneeds a request
the data needs a request or a data use agreement
Licence checked on 2026-10-10 at https://pmc.ncbi.nlm.nih.gov/articles/PMC11993632; derived test pairs shareable: unclear; commercial use: unclear. Sci Data 2025; the agreement text is not reproduced in the paper
UW-Madison GI tract MR-linac0.35 T MR-linac MR on different days, same patient~360 inter-day pairs (estimate)stomach, small and large bowel on every scan, supervised by a radiation oncologistboth imagesaccountCC BY-NC 4.0non-commercialneeds an account
the download needs an account
2D PNG slices with RLE masks, must be rebuilt into 3D
Vanderbilt laparoscopic-to-open liver deformation cohort (Heiselman 2018)preoperative CT liver model to intraoperative sparse surface data at two insufflation pressures and open conversion, same patient25 patients plus a mock laparoscopic phantomsparse tracked surface data; phantom subsurface targetsone imagehiddennot releasedaskleft out
its images or its truth are not released
Clinical part not public; the phantom line continues in the listed Image-to-Physical Liver Registration Sparse Data Challenge.request; the sparse data challenge supplies the public analogue
WAW-TACEmulti-phase CT before and after chemoembolisation, same patient233 patientstumour masks on one phase only; organs by TotalSegmentatorone imagedirectCC BY 4.0opentruth to make
no truth on both images; Two people place 15 landmarks (hepatic and portal vein branches, vertebrae, kidney poles) on pre- and post-TACE CT of 10 patients
organ truth from a programTwo people place 15 landmarks (hepatic and portal vein branches, vertebrae, kidney poles) on pre- and post-TACE CT of 10 patients; the TotalSegmentator organs on both are a weaker check.
WORDCT, inter-patient15016 organs including rectum and bladder, manualboth imagesrequestGPL-3.0 and academic research onlynon-commercialneeds a request
the data needs a request or a data use agreement
adds pelvic organs

Pelvis · 29

DatasetProblemPairsChecked againstAccessLicenceStatusNotesTo make the truth
AIDA CT pelvis landmarkspelvic CT, between patients (from TCIA CT Lymph Nodes and CT Colonography)90 CT15 anatomical landmarks and 5 bone masks per scanboth imagescontrolledAIDA license, controlled accessaskneeds a request
the data needs a request or a data use agreement
Inter-patient landmark truth; Swedish AIDA hub.
BTCV Multi-Atlas Labeling Beyond the Cranial Vault: Cervixradiotherapy planning CT of cervical cancer patients, between subjects (label overlap)30 labelled training subjects, 20 test subjects with hidden labels; 600 elastix registrations train to test suppliedbladder, uterus, rectum, small bowel delineated in clinical practice by a trained radiation oncologist; most patients carry four implanted markers around the cervixboth imagesaccountno named licence: 'access under the conditions of the data sponsors'; register on Synapse and agree to the licence termsaskneeds an account
the download needs an account
Erasmus MC. One planning CT per patient as far as the page says, so the markers give no intra-patient truth. Mix of prone (belly board) and supine; some organ contours missing. Licence checked on 2026-10-10 at https://repo-prod.prod.sagebase.org/repo/v1/entity/syn3193805/wiki/217753; derived test pairs shareable: unclear; commercial use: unclear. same challenge project as the Abdomen set
COBRA2026 (pelvic CBCT projections)Raw CBCT projections + planning CT deformably registered to daily CBCT, same patient867 cases, 6 centres (692/52/123)Organiser deformable CT-to-CBCT registration (not independent truth); acquisition geometryone imagedirectCC BY-NC 4.0non-commercialtruth to make
no truth on both images; threshold gold seeds on original CT and RTK CBCT, centroid = landmark, a person checks
Planning CT already warped to CBCT; unregistered CT availability unverified status change: the paper lists ct_original.mha, the planning CT without any registration, in every case, so CT-to-CBCT and CT-to-projection (2D/3D) registration can be scored. Metal implants and rectal balloons kept on purpose; prostate gold seeds likely but unverified.threshold gold seeds on original CT and RTK CBCT, centroid = landmark, a person checks; seeds also visible on raw projections, so known geometry gives 2D/3D marker truth; limits: seeds migrate slightly, metal streaks, test set held by challenge
CPTAC-UCEC-Tumor-Annotations (on CPTAC-UCEC)MR, CT of endometrial cancer, same patient across trial timepoints (baseline, on-treatment, follow-up)72 patients annotated; 17 with >=2 annotated studies; 9 lesions trackedNCI annotation project: up to 5 RECIST/PERCIST lesions per scan segmented, same lesions tracked over time (DICOM Tracking ID/UID) plus seed points; radiologists, US board-certified reviewboth imagesdirect: TCIA's NBIA API or Data Retriever manifest, no accountCC BY 4.0 (annotations and the CPTAC-UCEC images, TCIA pages)openin the corpuspriority 3
In the corpus as segmentation; 7 patients share a tracked lesion.
Sparse lesion correspondences (centroids/seed points with tracking IDs) and lesion masks on both scans; lesions change size so use as landmarks, not dense overlap. Few longitudinal cases.
Fedorov MR-TRUSMR to transrectal ultrasound, prostate, same patient3gland on both and fiducials (urethra base, verumontanum, apex, cysts)both imagesdirectCC BY-NC-SA 4.0non-commercialin the corpuspriority 3
Built: FedorovMRTRUS (track prostate-mrus-three-cases).
Gold Atlaspelvis MR and CT, same patient19contours by five expertsboth imagesrequestrestricted on Zenodoaskneeds a request
Haukeland rectumrectal contour meshes–contour meshes only, no imagesone imagedirectCC0opentruth to make
no truth on both images; Contour meshes only, no images: no usable registration truth.
Contour meshes only, no images: no usable registration truth.
Homologous Point Transformer prostate MR-histologyT2 MRI slice vs whole-mount H&E, same patient (slicing jig)Patients x slices (unverified count)Manually curated homologous control points on bothboth imagesdirectCC BY 4.0openwaits for a tool
2D MR slice to 2D whole-mount histology.
2D, 512x512
IMAG2 paediatric pelvis T2w and DWI (Necker, private)T2w MRI to DWI b0, same session, children with pelvic tumours or malformations80 patientship bones, sacrum, L5 and bladder labelled on T2w and b0 by an expert and checked by two moreboth imageshiddenProprietaryaskleft out
its images or its truth are not released
Bone labels on b0 transferred by a checked rigid registration. status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Multimodal DIR benchmark 2026).
Low-resolution prostate MR on MR-Linac (AIDA LESSER)repeated T2 MRI on an Elekta Unity MR-Linac, healthy volunteers, about 5 minutes apart, same subject (prostate, bladder filling, rectum motion)310 scans, 255 annotatedprostate (CTV), bladder and rectum on the initial and all high-resolution scans, single observerboth imagesrequestControlled access: free for use in legal and ethical medical diagnostics research via a data sharing agreement; AIDA BY licence only within AIDAaskneeds a request
the data needs a request or a data use agreement
Partly fills the bladder-filling gap (minutes apart, volunteers). Licence checked on 2026-10-10 at https://datahub.aida.scilifelab.se/10.23698/aida/lesser; derived test pairs shareable: unclear; commercial use: unclear. request e-mail from an institutional account (PI with PhD), agreement template on the AIDA site
LUND-PROBEMR and synthetic CT, prostate432contours; the synthetic CT is made from the MRboth imagescontrolledcontrolled accessaskneeds a request
the data needs a request or a data use agreement
circular pair
MEGRE (prostate gold fiducial MRI, AIDA)prostate multi-echo gradient-echo MRI (8 echoes) with 3 implanted gold seeds, one session per patientnone across sessions: 326 training + 40 test patients, one session eachcentre of mass of each gold seed placed by hand (several observers), confirmed on CT that is not releasedone imagerequestControlled access: free for use in legal and ethical medical diagnostics research via a data sharing agreement; AIDA BY licence only within AIDAaskneeds a request
the data needs a request or a data use agreement
Fails pairs: one image session, CT not included. Same group's LUND-PROBE (listed) has MR + seeds + synthetic CT. Licence checked on 2026-10-10 at https://datahub.aida.scilifelab.se/10.23698/aida/megre; derived test pairs shareable: unclear; commercial use: unclear. request template says publications must include the dataset authors in the author listonly useful if the matching planning CT is obtained on request: seed centroids on CT vs MEGRE give same-week multimodal point truth (3 per patient); limits: seeds migrate slightly, calcifications mimic seeds on MR
MR-guided prostate biopsy pre/intra-procedural MRI registration evaluation dataset (Fedorov 2012)pre-procedural endorectal-coil T2 MRI to intra-procedural T2 MRI during in-bore transperineal biopsy, same patient10 patientsgland, central and peripheral zone segmentations; 34 landmarks (3-4 per patient); 36 biopsy targets; 2 radiologistsboth imageshiddenunknownunknownleft out
its images or its truth are not released
The cited URL na-mic.org/publications/item/view/2147 returns 404; ask Fedorov (SPL). NA-MIC 2016 mentions 250+ internal cases.
MRgRT deformable dose accumulation comparison cases (Murr 2024, six centres)online adaptive MR-guided RT fractions: prostate (2), cervix, liver, lymph node, plus a biomechanically deformed gold-standard case; same patient6 cases, many fractionsgold-standard case: known biomechanical deformation; clinical cases: contours onlyboth imageshiddennot releasedaskleft out
its images or its truth are not released
Shared among the six centres only.request from Tübingen; the gold-standard idea (biomechanical warp of a real MR) gives exact truth
mu-RegPro (MR to TRUS registration for prostate, MICCAI 2023)Pre-operative prostate MR to 3D transrectal ultrasound, same patient; volumes are sliced into 2D TRUS frames for slice-to-volume work (DreamReg, SCoPE-Reg)100+ paired MR-TRUS (train and validation released)Expert-annotated corresponding landmarks and gland labels on both imagesboth imagesdirectCC BY-NC-SA 4.0non-commercialleft out
Already in the corpus as muRegPro (same train.zip); only its 12.7 MB val.zip is new and can join that entry.
Native task is 3D-3D; 2D TRUS-to-MR by slicing the TRUS volume (known slice pose)Slice the TRUS volume at recorded poses for exact 2D-to-3D truth; MR-TRUS landmarks give the cross-modal check
NaF PROSTATE18F-NaF PET/CT, some double baselines within 14 days, follow-up at 6 and 12 months9 patients, 44 studiesNoneone imagedirectCC BY 3.0opentruth to make
no truth on both images; PET and CT of one PET/CT scan are aligned by the scanner, so a known warp on the CT gives exact PET-CT truth
Fails truth; test-retest PET/CT.PET and CT of one PET/CT scan are aligned by the scanner, so a known warp on the CT gives exact PET-CT truth; double baselines within 14 days are near identity for bone, so two people place 15 vertebra and pelvis landmarks on all 9 patients for a rigid check.
Odense prostate MR-linac structure propagation cohort (Christiansen 2020)planning CT and MR to three daily T2 MR on a 1.5 T MR-linac, high-risk prostate, same patient12 patients x 3 fractionstarget and organ contours delineated on planning MR, CT and fraction MRboth imageshiddennot releasedaskleft out
its images or its truth are not released
Not public.request from Odense; gold seeds or calcifications as point truth if present
PENGWIN 2026 Task 3 (pelvic fracture reduction planning)rigid reassembly of pelvic fracture fragments (meshes and point clouds from CT) to their anatomical position, same patient170 clinical cases with fragments; 480 simulated subjects x 20 fracture variants per bonerigid reduction matrix per fragment: expert-made for clinical cases (method unverified), exact for simulated onesboth imagesdirectCC BY 4.0 NonCommercial (Zenodo)non-commercialwaits for a tool
fragment meshes and point clouds, not image pairs.
Shape/mesh registration, not image pairs; source CT is in Task 1/2 record 19732767 (500 cases, CC BY-NC-SA, 19.6 GB). Test phase hidden.simulated fractures already give exact truth; for the 170 clinical cases two surgeons plan reductions independently to measure spread
PI-CAI public set (includes PROSTATEx)prostate bpMRI: axial/sagittal/coronal T2w, high-b DWI, ADC of one exam1,500 exams (328 from PROSTATEx)csPCa lesion delineations by trained readers under expert radiologists, all 1,500 cases human since 2025both imagesdirectCC BY-NC 4.0non-commercialtruth to make
one lesion map per exam, not one per image; a known warp between contrasts would make the truth (the known-warp track's kind); prostate MR already covered.
DWI and ADC come from one acquisition (aligned by construction); T2w vs DWI moves slightly and DWI is distorted. Some lesion masks were drawn on ADC/DWI, others on T2w. Do not combine with PROSTATEx (duplicate cases).warp T2w by a known field and register ADC/DWI to it: truth exact up to EPI distortion; lesion masks warped along give a lesion-overlap check
Princess Margaret prostate MR-linac fractions (Malkov 2025)1.5 T MR-linac T2 MR: reference vs adapt, verify and beam-on MR of each fraction, same patient25 patients, 375 fraction images, 1,875 DIRsCTV, bladder, rectum contoured on every image by a trained therapist, reviewed by an oncologist and a physicistboth imageshiddendata availability: 'will not be publicly available'askleft out
its images or its truth are not released
Intra- and inter-fraction prostate MR with large bladder changes; not public.request from Princess Margaret; contours exist, landmarks (seeds, calcifications) by two people would add point truth
ProMotion LMU prostate motionprostate position traces from 4D transperineal ultrasound during radiotherapy3,423 fractions, 191 patients (2024 edition)positions from the ultrasound system's own registrationdirectCC BY 4.0openleft out
no images: prostate position traces as CSV only (2019, 2022 and 2024 editions)
Fails: CSV positions only, no images (2019, 2022, 2024 editions).not possible without images
Prostate Fused-MRI-Pathologyin vivo 3T prostate MRI (T1, T2, DWI, DCE) to whole-mount H&E histology of the prostatectomy specimen, same subject28 patientsMRI-histology slice correspondences, pathologist cancer annotations on histology, cancer extent mapped onto MRIboth imagesdirectCC BY 3.0openwaits for a tool
2D histology to MRI slices, and the MRI cancer maps come from the creators' registration.
MRI cancer maps come from the creators' own deformable registration (circular for scoring); pseudo whole-mount sections stitched from quarters; 2D-to-3D slice problem. status change: Round 4 count: listed in 2 of the 38 round-4 surveys and benchmark papers (Hammoudeh 2023 survey; WSI registration review 2025).
PROSTATE-MRI (NCI, TCIA)Endorectal 3T prostate MRI and whole-mount prostatectomy histology cut in an MRI-based mould (MR/histology)26 patientsSlice correspondence by construction (specimen sectioned in the MRI plane using a patient-specific mould); no landmarks seenone imagedirectCC BY 3.0opentruth to make
no truth on both images; The mould fixes slice correspondence
Truth limited to mould-based slice matching.The mould fixes slice correspondence; two people place 8-10 landmarks (urethra, ejaculatory ducts, nodules, capsule points) on each MRI slice and its whole-mount section for 10 patients.
Prostate158T2, ADC, DWI prostate MR, same patient158gland zones and lesionsdirectnone stated (Zenodo records 6481141/6592345 have no licence; repo MIT covers code)unknownwaits for a toolpriority 2
Waits for a tool: the pairs (a provider test study with a training study) would hold Prostate158's segmentation images, and fidopt's 'from' derives cases from a registration set's pairs, not pairs from a segmentation set.
Licence checked on 2026-10-10 at https://zenodo.org/api/records/6481141; derived test pairs shareable: unclear; commercial use: unclear. README: training data 'freely available'; paper data statement not reachable (Elsevier blocked)Draw the gland and zones on ADC for 10 test cases (T2 is labelled): contours on both; DWI has EPI distortion, so a known warp on T2 gives only near-exact multi-contrast truth.
RAPSODI / Stanford radiology-pathology cohortprostate MRI to whole-mount histopathology (3D-printed mold), same subject157 subjects in RAPSODI, 907 slice pairs; 152 in later Stanford work257 landmarks visible on both MRI and histology for 12 subjects (BPH nodules, ejaculatory ducts), prostate and urethra outlinesboth imageshiddenunknownleft out
its images or its truth are not released
Only code is public (Slicer-RadPathFusion, ProsRegNet); data not released. Ask the authors.
Robarts MR to TRUS prostate cases with intrinsic fiducials (Sun 2015)T2 MR to 3D TRUS of the prostate, same patient20 patientsmanually identified corresponding intrinsic fiducials in the whole gland and peripheral zoneboth imageshiddennot releasedaskleft out
its images or its truth are not released
Not public.request from Western University (Robarts)
SynthRAD2023pelvic CBCT and planning CT (task 2), pelvic MR and CT (task 1), same patient; prostate, rectal and cervical RT patients (brain subsets too)pelvis: 270 CBCT-CT and 270 MR-CT pairs (180 training each released first; validation and test later); brain the samenone; organisers aligned each pair rigidly with elastixone imagedirectCC BY-NC 4.0non-commercialtruth to make
no truth on both images; in prostate cases, threshold the gold seeds on CT and CBCT (signal voids on MR), centroid per seed = landmark, 3 seeds per patient, a person checks the match
Resampled to 1x1x2.5 mm and cropped. Dutch prostate patients usually carry 3 gold seeds; whether seeds show in the released pelvis images is unverified (paper mentions hip implants and calcifications only).in prostate cases, threshold the gold seeds on CT and CBCT (signal voids on MR), centroid per seed = landmark, 3 seeds per patient, a person checks the match; limits: 2.5 mm slices blur centroids, seeds can migrate a little, prostate deforms between days
UCL MR to TRUS prostate fusion cases (Hu 2012)T2 MR to 3D transrectal ultrasound of the prostate, same patient8 patientscorresponding anatomical landmarks on MR and TRUSboth imageshiddennot releasedaskleft out
its images or its truth are not released
Not public; mu-RegPro (listed) is the public successor from the same group.request from UCL
UCL SmartTarget MR-TRUS prostate pairs (Hu et al.)T2 MRI to 3D transrectal ultrasound of the prostate, same patient108 pairs from 76 patientsgland labels and ad hoc landmarks (cysts, calcifications) by experts on bothboth imageshiddenprivate (UCL trials)unknownleft out
its public successor, µ-RegPro, is in the corpus
Fails access. The public successor is mu-RegPro (already in the corpus).use muRegPro (in the corpus) for public MR-TRUS

Whole body · 29

DatasetProblemPairsChecked againstAccessLicenceStatusNotesTo make the truth
AHOD0831-Tumor-Annotations (on AHOD0831 (COG Hodgkin lymphoma))CT and FDG PET/CT of pediatric Hodgkin lymphoma, same patient across trial timepoints (baseline, on-treatment, follow-up)165 patients; 164 with >=2 annotated studies; 517 lesions trackedNCI annotation project: up to 5 RECIST/PERCIST lesions per scan segmented, same lesions tracked over time (DICOM Tracking ID/UID) plus seed points; radiologists, US board-certified reviewboth imagescontrolledannotations CC BY 4.0; images NCTN/NCORP licence, now NIH Controlled Data Access Policyaskneeds a request
the data needs a request or a data use agreement
Sparse lesion correspondences (centroids/seed points with tracking IDs) and lesion masks on both scans; lesions change size so use as landmarks, not dense overlap. NCTN trial images: TCIA lists them as "Unavailable" since July 2025 (moving to NCI CTDC, dbGaP request, trial-specific PHS); annotations themselves are public.
AIHub radiotherapy patients CBCT and EMR (Korea, dataset 71514)CBCT of radiotherapy patients fused with pathology and treatment records; planning CT unverifiedunverifiedunverifiedcontrolledAIHub terms; online safe zone, Korean nationals onlyaskneeds a request
the data needs a request or a data use agreement
found on AIHub; built for prognosis, not registrationif several fractions exist: bony landmarks and fiducials across CBCTs, checked by a person
BodyParts3D / Anatomographysurface meshes of one reference bodynonepart namesone imagedirectCC BY-SA 2.1 JP (unverified)opentruth to make
no truth on both images; One reference body: no pair
Shape set; only point-cloud registration. Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Point-cloud shape survey 2025).One reference body: no pair; useful only as a shape template for a known deformation.
CALGB50303-Tumor-Annotations (on CALGB50303 (DLBCL lymphoma))CT and FDG PET/CT of lymphoma, same patient across trial timepoints (baseline, on-treatment, follow-up)152 patients annotated; 129 with >=2 annotated studies; 270 lesions trackedNCI annotation project: up to 5 RECIST/PERCIST lesions per scan segmented, same lesions tracked over time (DICOM Tracking ID/UID) plus seed points; radiologists, US board-certified reviewboth imagescontrolledannotations CC BY 4.0; images NCTN/NCORP licence, now NIH Controlled Data Access Policyaskneeds a request
the data needs a request or a data use agreement
Sparse lesion correspondences (centroids/seed points with tracking IDs) and lesion masks on both scans; lesions change size so use as landmarks, not dense overlap. NCTN trial images: TCIA lists them as "Unavailable" since July 2025 (moving to NCI CTDC, dbGaP request, trial-specific PHS); annotations themselves are public.
Cancer Moonshot Biobank (CMB-AML, CRC, GEC, LCA, MEL, MML, OV, BRCA, PCA)Clinical CT, MR, PET/CT, US over the course of treatment, same patient, plus WSIse.g. CMB-LCA: 129 patients, 89 with >=2 studiesNoneone imagedirectCC BY 4.0; head images NIH Controlled Data Access Policy (separate dbGaP phs002192)opentruth to make
no truth on both images; Pick CT follow-up pairs (e.g. CMB-LCA)
Fails truth; longitudinal multi-modal clinical imaging.Pick CT follow-up pairs (e.g. CMB-LCA); two people place 20 landmarks on 10 pairs, TotalSegmentator labels on both as a weaker check; PET/CT studies give exact PET-CT truth under a known warp.
DEEP-PSMA (MICCAI 2025)PSMA PET/CT and FDG PET/CT of same patient, separate sessions100 train patients (both tracers), 10 val, 40 testTotalSegmentator organs on both CTs + expert lesion (TTB) labels on both PETs; coarse rigid alignment providedboth imagesaccountCC BY-NC 4.0non-commercialneeds an account
the download needs an account
Inter-session whole-body CT-CT and cross-tracer PET pairs with labels on both Licence checked on 2026-10-10 at https://zenodo.org/api/records/15281784; derived test pairs shareable: yes; commercial use: no. training data released on Zenodo 15281784 (24 Nov 2025 news on the challenge page)
DeepLesion tracking (DLS)CT follow-up sub-volumes, same patient3,891 lesion pairsmatched lesion centres, boxes and RECIST measurementsboth imagesdirectannotations CC BY-SA 4.0; DeepLesion images unverifiedunknownleft outPartial slabs, not whole volumes. status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Foundation models review 2026).
DIR-DB radiotherapy re-planning database (Kanazawa University)repeat radiotherapy planning CTs and setup CBCTs, same patient (head and neck, thorax, abdomen, pelvis, prostate brachytherapy)48 patients with 2 to 4 plans; CBCT for 17clinical contours on each planning CTboth imagescontrolledaccess for users in Japan with IRB approvalaskneeds a request
the data needs a request or a data use agreement
URL is the describing paper (Takemura et al. 2020); the database site was not checked; contours change when targets are redefinedtwo people place 15 bony and soft-tissue landmarks on 10 plan pairs; contour overlap already available
Duke chest-abdomen-pelvis CT with organ labels (XCAT-to-CT, Segars 2013)non-contrast CT to XCAT phantom, atlas to patient50 CT scansmanual organ labels (16 structures) by students, corrected by a radiologistboth imageshiddenprivate (Duke)unknownleft out
its images or its truth are not released
Fails access.
EAY131-Tumor-Annotations (on EAY131 (NCI-MATCH))CT, some PET/CT and MR, many tumour types, same patient across trial timepoints (baseline, on-treatment, follow-up)2,407 patients annotated; 73 with >=2 annotated studies; 61 lesions tracked (27 patients)NCI annotation project: up to 5 RECIST/PERCIST lesions per scan segmented, same lesions tracked over time (DICOM Tracking ID/UID) plus seed points; radiologists, US board-certified reviewboth imagesdirect: TCIA's NBIA API or Data Retriever manifest, no accountCC BY 4.0 (annotations and EAY131 images, TCIA pages)openleft outpriority 3
TCIA warns the EAY131 volumes have missing slices (found Summer 2026, fix pending); build as tumours-same once corrected.
Sparse lesion correspondences (centroids/seed points with tracking IDs) and lesion masks on both scans; lesions change size so use as landmarks, not dense overlap. Images public (CC BY 4.0) but few longitudinal annotations.
ENHANCE.PET whole/total-body FDG PET/CT with 130 CT regionswhole- and total-body FDG PET/CT, healthy and diseased, between subjects (PET and CT same session)1,683 PET/CT studies130 CT regions segmented by software then corrected by physicians in trainingboth imagesdirectCC BY 4.0 (Careggi, Leipzig); CC BY-NC 4.0 (autoPET part)non-commercialto build
The Sci Data 2026 article of ENHANCE.PET 1.6k, open on AWS (s3://enhance-pet-1-6k) without a request: no request or agreement is needed. Its Leipzig and Careggi CTs and their 127 corrected structures are built for segmentation as EnhancePETBodyCT, and Leipzig CT to PET as ENHANCEPET_CTPET_Syn (other patients); an inter-subject registration set would use other patients of the same release.
Sci Data 2026. Access route not found on the article page (unverified); TotalSegmentator, SAROS and WORD sources are open.two people place 20 skeletal landmarks on 20 subjects for inter-subject TRE
Fetal pose keypoints in EPI time series (MIT, Xu 2019)Fetal BOLD EPI MR volumes over time (25-35 weeks); same subject, motion between frames1705 labelled volumes from 70 pregnancies (pairs within a series)15 manual keypoints (joints, bladder, eyes) per labelled volumeboth imageshiddennot released (no public link found)unknownleft out
its images or its truth are not released
Would be a strong intra-subject motion set if released; ask MIT (Golland/Adalsteinsson).
Healthy-Total-Body-CTslow-dose total-body CT from uEXPLORER PET/CT at several time points the same day, same subject30 subjects (15 x 3 time points over 3 h, 15 x 6 over 12 h)MOOSE automatic segmentations of 37 structures at the 90-minute time point onlyone imagecontrolledCT images NIH controlled access; segmentations and clinical CC BY 4.0askneeds a request
the data needs a request or a data use agreement
Truth on one time point; silver labels could be made by running the same tool on the others.Run MOOSE on every time point for labels on both (a weaker check); two people place 20 skeletal and vessel landmarks on 10 same-day pairs.
J-MID (Japan Medical Image Database)CT and MRI with reports from 7 Japanese university hospitalsunverifiedannotations being made for AI (unverified)controlledJapan Radiological Society agreement (unverified)askneeds a request
the data needs a request or a data use agreement
Japan-only access; no registration truth.two radiologists mark landmarks on 20 follow-up pairs
JHU chest-abdomen-pelvis CT (TransMorph)CT, between patientsunverifiedorgan labelshiddennot publicunknownleft out
its images or its truth are not released
Listed by Azad 2023 (TransMorph); private. Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Azad 2023 ViT review).use TotalSegmentator-labelled public CT (e.g. AbdomenCT-1K)
Lesion Tracking Dataset for 3D TBP (Huang et al. 2024, on 3DBodyTex)Textured 3D total-body scans (photographic texture on surface meshes) of the same person in two poses; match skin lesions across scans198 subjects, about 25,000 annotated lesion pairs (JHU); Skin3D adds 25,000+ lesion boxes with multi-annotator test setManually annotated corresponding lesions on both meshes (triangle id and xyz), ordered row-to-rowboth imagesrequestannotations: no licence file; meshes: 3DBodyTex.v1 signed agreement (Univ. Luxembourg)askneeds a request
the data needs a request or a data use agreement
Surface-to-surface with texture, not photo-to-volume; both scans same session in different poses, not longitudinal Licence checked on 2026-10-10 at https://github.com/weilunhuang-jhu/LesionTrackingDatasetTBP3D; derived test pairs shareable: unclear; commercial use: unclear. skin3d README: 'You need to sign an agreement in order to access the data'
Mayo proton therapy longitudinal CT (private)planning CT to repeat CT in proton therapy, many sites1,222 pairs, 553 patientsclinical CTV and OAR contours and dose (no landmarks)hiddenInstitutionalaskleft out
its images or its truth are not released
Largest DIR cohort reported; not shared.two people place landmarks on 20 pairs
MedPointSmedical point clouds from segmentationsbetween shapespart labelsdirectMIT tag on the HF datasets (wlsdzyzl/MedPointS-cls/-cpl/-seg), but the point clouds are built from MedShapeNet shapes whose sources carry mixed licences (some NC/ND)askwaits for a tool
point clouds only (built from MedShapeNet), 1,020 per-patient multi-structure samples in its segmentation set: waits for mesh and point-cloud scoring
Named in the point-cloud survey; page not found. Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Point-cloud shape survey 2025). Licence checked on 2026-10-10 at https://flemme-docs.readthedocs.io/en/latest/medpoints.html; derived test pairs shareable: unclear; commercial use: unclear. MIT is the creators' tag; upstream source licences may still apply per shapePoint clouds derived from segmentations, no image pairs: no registration truth possible from it alone.
MedShapeNet100,000+ anatomical shapes from public segmentationsbetween shapesshape labelsdirectper-source licences listed by the maintainers: CC BY 4.0 (e.g. TotalSegmentator, VerSe, SkullBreak, AMOS), CC0 (CT-ORG), but also CC BY-NC-SA/NC-ND (EMIDEC, LNDb, 3DTeethSeg), TCIA Restricted (GLISRT), HCP terms, and blanksaskwaits for a tool
meshes only, from public segmentations; structure names are the only correspondence: waits for mesh scoring; mixed source licences
Shape set for point-cloud registration; no images. Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Point-cloud shape survey 2025). Licence checked on 2026-10-10 at https://github.com/GLARKI/MedShapeNet2.0; derived test pairs shareable: unclear; commercial use: unclear. mixed: only the CC BY/CC0 sources are open; download site medshapenet.ikim.nrw shows no licenceShapes, not images: only a known deformation of a shape; no image registration truth.
QUADRA_HC total-body FDG PET/CT test-retest (Zenodo 16686025)total-body FDG PET/CT (Biograph Vision Quadra), 48 healthy adults, test and retest scans, same subject48 test-retest pairs (CT and PET)MOOSE automatic organ segmentations, sanity-checked (presence on both sessions unverified)direct: Zenodo, one 20.2 GB zip whose members read by HTTP ranges (listed and NIfTI headers read that way)CC BY 4.0 on the Zenodo record and in the Scientific Data descriptoropenin the corpuspriority 2
48 healthy adults scanned twice on a Biograph Vision Quadra 38 +/- 10 days apart (22 to 98), arms down, low-dose CT of 512x512 at 1.52 x 1.52 x 2 mm (531 slices for HC_003) with PET; all 48 have both sessions. The truth is MOOSE 3.0.13's 135 automatic structures on each CT in eight multi-label files, which the authors did not verify one by one (sanity checks of laterality, presence and gross shape only), so it ranks apart as automatic. No corpus dataset holds these subjects; whole-body same-patient CT in the corpus is LongitudinalCT (non-commercial).
Open whole-body CT rescan with organ labels; test-retest interval unverified. Different from Healthy-Total-Body-CTs (listed).MOOSE or TotalSegmentator labels on both (weaker); two people place 20 skeletal and vessel landmarks on 10 pairs; PET/CT hybrid gives exact multimodal pairs under a known warp
RadImageNetCT, MR, US slices with diagnostic labelsnoneimage labelsone imagerequestunverifiednon-commercialneeds a request
the data needs a request or a data use agreement
Fails pairs: one image per subject, no repeat or second modality. Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Foundation models review 2026).Unpaired slices with diagnostic labels: no usable truth.
RUMC whole-body longitudinal CT (TotalRegistrator)thorax-abdomen-pelvis CT, same patient over time591 train, 104 evaluation pairsTotalSegmentator labels (automatic)both imageshiddenin-houseunknownleft out
its images or its truth are not released
Training and test data of TotalRegistrator 2025; private. Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (TotalRegistrator 2025).
Skin3D lesion annotations on 3DBodyTex3D textured total-body meshes, same subject two poses, lesion boxes in 2D texture images and lesion IDs across scans10 subjects with longitudinal lesion IDs (~20 pairs each); 25,000+ lesion boxes overallmanual lesion boxes (multiple annotators on test) and IDs linking the same lesion across two scansboth imagesrequestrepo licence NOASSERTION; meshes need the signed 3DBodyTex.v1 agreementaskneeds a request
the data needs a request or a data use agreement
Superseded in size by the Huang 2024 lesion-pair set on the same meshes Licence checked on 2026-10-10 at https://github.com/sfu-mial/skin3d; derived test pairs shareable: unclear; commercial use: unclear. README: sign an agreement at cvi2.uni.lu/datasetsuse the Huang 2024 pairs
Slicer Registration Case Library (NA-MIC RegLib)about 30 real registration problems: brain tumour and MS follow-up, DTI to T2, breast and prostate MR pre/post, abdominal MR-CT, PET/CT follow-up, knee, vertebra, mouse brain, wood micro-CTone pair or a few per case (case IDs 1 to 50, about 30 published)mostly none: example fiducials in a few cases (C15, C17, C48, C49, C50) and the authors' solution transformsdirectdata licence not stated; repository text Apache-2.0, Slicer wiki case pages CC BY-SA (page text)unknownleft out
its download links point to the retired slicer.kitware.com MIDAS server, which answers 404, and no mirror was found
Data were served from NA-MIC MIDAS; whether the .mrb links still work is unverified. Fails the truth criterion except for a few fiducial sets. Licence checked on 2026-10-10 at https://github.com/Slicer/SlicerRegistrationLibrary; derived test pairs shareable: unclear; commercial use: unclear. MRB files linked to the retired slicer.kitware.com/midas3 servertwo people place 10 to 20 landmarks per case; for C17/C48 abdominal MR-CT the example fiducials can be completed
TotalSegmentator CTCT, inter-patient1,228 scans117 classes from a network, corrected in a loopboth imagesdirectCC BY 4.0openleft outTruth largely made by a program.
UDPET (Ultra-low dose PET challenge 2022-2024)total-body FDG-PET (Siemens Vision Quadra, United Imaging uEXPLORER), full dose and simulated low dose from list mode1,447 subjects, dose reduction 4-100xnone; low-dose images subsampled from the same list mode, perfectly alignedrequestsigned Data Transfer Agreement (Univ. Bern / Ruijin): academic purposes in an approved research plan only; no commercial use; no distribution or disclosure to third parties; team members bound toonon-commercialneeds a request
the data needs a request or a data use agreement
Monomodal, no CT provided. Exact alignment across noise levels: a clean known-warp test with controllable noise. Licence checked on 2026-10-10 at https://drive.google.com/file/d/1-dGYEB6lWSJeY8eI5XFtT3MtJGOqXAX1/view; derived test pairs shareable: unclear; commercial use: no. DTA linked from udpet-challenge.github.io/dataset.html; manuscripts must be sent to UNIBERN before submissionwarp the full-dose PET by a known field and register the low-dose PET to it: exact truth with noise level as a difficulty knob
UK Biobank, NAKO, raw OAIlarge cohortsvery manymostly automaticrequestaccess only through each resource's application and access/data use agreement (UK Biobank site blocks scripts; terms not re-read)askneeds a request
the data needs a request or a data use agreement
status change: usage counted from papers: UKBB Dixon fat/water pairs are aligned by acquisition, a free multimodal truth under a known warp. status change: Round 4 count: listed in 8 of the 38 round-4 surveys and benchmark papers (Bharati 2022 review; Zou 2022 review (Front Oncol); Hammoudeh 2023 survey; Chen X 2021 review (Prog Biomed Eng); Foundation models review 2026; Point-cloud shape survey 2025; uniGradICON 2024; multiGradICON 2024). Licence checked on 2026-10-10 at https://www.ukbiobank.ac.uk/; derived test pairs shareable: no; commercial use: unclear. none of the three allows passing data to people outside the approved project; NAKO and OAI (NDA) have separate DUAsNeeds an application; on a pulled subset of repeat-visit pairs, two people place 20 landmarks on 10 pairs, TotalSegmentator labels on both as a weaker check; data cannot be redistributed.
VISCERAL Anatomy3 (and Gold Corpus)whole-body CT, contrast thorax-abdomen CT, whole-body T1 MR, contrast fat-saturated abdominal T1 MR; between subjects120 volumes, 30 per modality (CADS re-hosts 40 CT; MR availability unverified)up to 20 organs and up to 53 named landmarks per volume by radiologists (VISCERAL chapter: 1,920 structures, 4,860 landmarks; the CADS card says 1,295 and 1,760)both imagesrequest'Academic research use only; Non-commercial use only; Internal use only (not transferable, no sublicensing, no redistribution of original images)'; original images via an application form, labels opennon-commercialneeds a request
the data needs a request or a data use agreement
visceral.eu is down; CADS says it obtained renewed permission. Whether any patient has both CT and MR is unverified. Landmarks are the same named points across subjects, so they give inter-subject point truth.
Whole-body MRI of neurofibromatosis patients (Synapse, MGH)whole-body STIR and T1 MRI of the same patient, NF1, NF2 and schwannomatosis48 patients, 96 stackstumour annotations unverifiedaccountunverifiedunknownneeds an account
the download needs an account
same-session contrasts, so mostly a multicontrast test with small motionvertebral and joint landmarks on both contrasts of 10 patients by two people

Brain · 195

DatasetProblemPairsChecked againstAccessLicenceStatusNotesTo make the truth
250 µm in vivo 7T human phantom (Lüsebrink 2017, 2021)One healthy adult, 7T: eight 250 µm T1 MPRAGE runs over five sessions, 150 µm TOF, 330 µm QSM, 450 µm T2, back-to-back 700 µm T1, DTI, >120 T1/PD volumes over ten years; same subjectMany same-subject pairs (between sessions, between contrasts, over 10 years)None; authors registered all contrasts to the 250 µm T1 (registrations archive)one imagedirectCC0 1.0 (OVGU repository; Dryad 10.5061/dryad.38s74 CC0)opentruth to make
no truth on both images; Rigid truth by construction between same-day runs (head is rigid): check non-rigid methods return near-identity
Prospective motion correction; brain shape stable over sessions, so inter-session truth is close to rigid.Rigid truth by construction between same-day runs (head is rigid): check non-rigid methods return near-identity; landmarks (AFIDs) placed twice on 250 µm and 700 µm T1 to test multi-resolution and long-term pairs
3T and 7T paired T1w/T2w MRI (UNC, 2023)3T and 7T T1w and T2w, same healthy subject10 subjects (unverified)tissue annotation maps and cortical surfaces; images pre-aligned by a programboth imagesdirectCC BY 4.0opentruth to make
10 subjects already aligned by a program for synthesis, so only a known warp would make a truth; tissue maps not stated as manual.
Different from the Chu 2025 hippocampal-subfield set already listed.two people place 20 landmarks on 10 subjects
7T multimodal ex vivo human brains (6 Chinese donors)ex vivo 7T structural, diffusion, QSM, T1 and T2 maps of the same brains; templates across brains6 brainssegmentation labels; within-brain maps share a sessiondirectunverified (figshare)unknownto buildpriority 2
Inter-subject brain with manual labels is covered (Mindboggle-101, UltraCortex); adds ex vivo 7T, 6 hemispheres (15 pairs); who drew the label maps is not stated
Sci Data 2025. Small; between-brain registration with labels.two people place 20 landmarks on all 6 brains
ABCD Study (release 6.0)T1, T2, DWI, fMRI of children 9-16 at 21 sites, same subject every 2 years11,868 participants; imaging at baseline, 2 and 4 years and about half of the 6-year visitFreeSurfer outputs (automatic)both imagescontrolledNBDC Data Use Certificationaskneeds a request
the data needs a request or a data use agreement
Largest longitudinal brain pool; growth plus scanner changes; no redistribution. status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (multiGradICON 2024).two raters place the 32 AFIDs fiducials (published protocol) on both scans of 20 pairs; SynthSeg labels on both
ABIDE I/IIT1 MR and rsfMRI, ~2,200 subjects (autism and controls)Inter-subject onlyNone manualdirectCC BY-NC-SA (INDI)non-commercialtruth to make
no truth on both images; Inter-subject only: AutoAFIDs places the 32 AFIDs on 20 subjects and a person checks them
Fails truth. status change: usage counted from papers: training pool only. status change: Round 4 count: listed in 9 of the 38 round-4 surveys and benchmark papers (Fu 2020 review (PMB); Boveiri 2020 review; Bharati 2022 review; Zou 2022 review (Front Oncol); Hammoudeh 2023 survey; Chen X 2021 review (Prog Biomed Eng); Chen 2025 survey (MedIA); Han 2022 GAN registration review; LUMIR 2025).Inter-subject only: AutoAFIDs places the 32 AFIDs on 20 subjects and a person checks them; SynthSeg labels on both as a weaker check.
ACNS0332-Tumor-Annotations (on ACNS0332 (COG medulloblastoma/PNET))MR (few CT) of pediatric brain tumours, pre-op to relapse, same patient across trial timepoints (baseline, on-treatment, follow-up)85 patients; 76 with >=2 annotated studies; 116 lesions trackedNCI annotation project: up to 5 RECIST/PERCIST lesions per scan segmented, same lesions tracked over time (DICOM Tracking ID/UID) plus seed points; radiologists, US board-certified reviewboth imagescontrolledannotations CC BY 4.0; images NCTN/NCORP licence, now NIH Controlled Data Access Policyaskneeds a request
the data needs a request or a data use agreement
Sparse lesion correspondences (centroids/seed points with tracking IDs) and lesion masks on both scans; lesions change size so use as landmarks, not dense overlap. NCTN trial images: TCIA lists them as "Unavailable" since July 2025 (moving to NCI CTDC, dbGaP request, trial-specific PHS); annotations themselves are public.
ACRIN-DSC-MR-Brain (ACRIN 6677 / RTOG 0625)Recurrent glioblastoma MRI at baseline and post-treatment107 evaluable patients with baseline + >=1 follow-upCentral neuroradiologist 2D/3D enhancement and FLAIR volume measurements only (no masks seen)one imagecontrolledNIH Controlled Data Access Policyaskneeds a request
the data needs a request or a data use agreement
Fails truth (no masks or landmarks). Images moved to dbGaP controlled access (phs004225, head images) in July 2025.Two raters place the 32 AFIDs, skipping those in the tumour, on 10 baseline/follow-up pairs; SynthSeg labels on both as a weaker check.
ACRIN-FMISO-Brain (ACRIN 6684)Glioblastoma: MRI, 18F-FMISO PET and low-dose CT, before and after treatment (PET/MR)45 patientsMR enhancing-tumour masks and PET hypoxia masks; hypoxia masks built after registering FLAIR to PETone imagecontrolledNIH Controlled Data Access Policy (images); CC BY 3.0 (clinical)askneeds a request
the data needs a request or a data use agreement
Masks are different structures and derived through registration: weak check. Images moved to dbGaP controlled access (phs004225, head images) in July 2025.PET and low-dose CT of the PET/CT are aligned by the scanner, so a known warp gives exact PET-CT truth; for MR to CT, two people place 15 skull and ventricle landmarks on 10 patients. Skip the hypoxia masks (made by registration).
ADHD-200T1 brain MRI of children with and without ADHD, between subjects949 scans used by RCN; about 970 subjectsnone manual; papers use FreeSurfer labelsdirectINDI sharing terms, non-commercial (as for ABIDE; not re-read for ADHD-200)non-commercialtruth to make
no truth on both images; SynthSeg or FreeSurfer labels on both, a weaker check
Fails truth. Training pool only in every paper seen.SynthSeg or FreeSurfer labels on both, a weaker check; two raters place AFIDs on 20 subjects for point truth
ADNILongitudinal T1 MR (with back-to-back repeat scans), AD/MCI/controlsThousands of longitudinal and repeat pairsNone manual except the HarP subsetone imagecontrolledADNI DUA, no redistributionaskneeds a request
the data needs a request or a data use agreement
Fails access. status change: usage counted from papers: mostly a training pool or a zero-shot target; ICNet added 5 ventricle landmarks per scan by a radiologist (not released). status change: Round 4 count: listed in 13 of the 38 round-4 surveys and benchmark papers (Haskins 2020 survey; Fu 2020 review (PMB); Boveiri 2020 review; Bharati 2022 review; Zou 2022 review (Front Oncol); Hammoudeh 2023 survey; Chen X 2021 review (Prog Biomed Eng); Abbasi 2022 scoping review; Chen 2025 survey (MedIA); Han 2022 GAN registration review; LUMIR 2025; rethink-reg 2024; ATOM 2024).Back-to-back repeats as near-identity, inverse and transitivity checks; two raters place AFIDs on 10 longitudinal pairs. No redistribution, so each user pulls the test images.
Aerobic Glycolysis Imaging PET-MRI (Ruijin, OpenNeuro ds004054)simultaneous FDG-PET and MRI (T1w, T2 map, T2* map, ASL CBF) on Biograph mMR; temporal lobe epilepsy and controls42 subjects (24 patients, 18 controls)nonedirectCC0opentruth to make
no truth on both images; warp the T2 map by a known field and register PET, T1w or T2* map to it: exact truth if the volumes are still in scanner space (check first)
Several quantitative MR maps plus PET from one simultaneous session: rich multimodal set. Small size suggests processed volumes; native vs standard space not checked.warp the T2 map by a known field and register PET, T1w or T2* map to it: exact truth if the volumes are still in scanner space (check first)
AFIDs on OASIS-1 and HCPT1 MR, inter-subjectmany32 expert fiducials a brainboth imagesaccountonly the SNSX and LHSCPD placements are CC BY 4.0 on OpenNeuro (already in the corpus)askneeds an accountstatus change: the AFIDs protocol is now the point truth of the main brain registration challenge (LUMIR), not only a fiducial-placement study. status change: Round 4 count: listed in 2 of the 38 round-4 surveys and benchmark papers (Learn2Reg 2024 overview; LUMIR 2025).
AHEAD (in vivo 7T)7T MP2RAGEME quantitative maps (R1, R2*, QSM...), 105 adults 19-80 y105 subjectsManual STN/SN/RN/GPi/GPe masks were drawn but only MNI probability maps are releasedone imagedirectCC BY 4.0 (figshare)opentruth to make
no truth on both images; R1, R2* and QSM maps come from one MP2RAGEME acquisition and are aligned by construction: a known warp gives exact multi-contrast truth
Fails truth: individual manual masks not shared (readme).R1, R2* and QSM maps come from one MP2RAGEME acquisition and are aligned by construction: a known warp gives exact multi-contrast truth; inter-subject pairs get AFIDs by two raters on 10 pairs.
AHEAD post-mortem brains (Alkemade 2022)Whole post-mortem brains: 7T qMRI (200-400 µm) vs blockface vs 200 µm sections, 5 stainsSeveral brains (7 planned), many sections eachAuthors' co-registration in blockface spaceone imagedirectCC BY 4.0 (UvA figshare records)opentruth to make
no truth on both images; Two people place 10 landmarks (vessels, sulcal points, tissue edges) on 10 section/blockface pairs and on blockface vs MRI
Host and licence not confirmed (likely UvA figshare/DataverseNL). Licence checked on 2026-10-10 at https://api.figshare.com/v2/articles/16844500; derived test pairs shareable: yes; commercial use: yes. paper: datasets 'shared without any restrictions'; stains, 3D reconstructions, parcellations on figshare DOIs 10.21942/uva.*Two people place 10 landmarks (vessels, sulcal points, tissue edges) on 10 section/blockface pairs and on blockface vs MRI; the authors' co-registration is a reference only.
AIBL (Australian Imaging, Biomarkers and Lifestyle)T1 MR and PiB PET, ageing and Alzheimer's, same subject at baseline, 18 and 36 monthsimaging subset released at 3 time points; subject count unverifiednoneone imagecontrolledAIBL Data Use Agreement via LONI IDA applicationaskneeds a request
the data needs a request or a data use agreement
Same access route as ADNI; no redistribution.two raters place the 32 AFIDs fiducials (published protocol) on both scans of 20 pairs; hippocampus traced by a person (HarP protocol) on both visits
AIHub brain metastasis diagnosis and treatment imaging (Korea, dataset 517)contrast brain MRI with brain CT and MRI, metastasis location and sizeunverifiedmetastasis annotations (on which images unverified)controlledAIHub terms; online safe zone, Korean nationals onlyaskneeds a request
the data needs a request or a data use agreement
found on AIHubAFIDs on CT and MRI of 10 patients by two raters
AIHub Parkinson's multimodal MR and PET (Korea, dataset 516)MR and PET DICOM of Parkinson's and atypical parkinsonism patients, same patientunverifiedpolygon or circle annotations of 8 anatomical regions (on MR, PET or both unverified)controlledAIHub terms; online safe zone, Korean nationals onlyaskneeds a request
the data needs a request or a data use agreement
found on AIHubif regions are drawn on both MR and PET: region overlap; else striatum and midbrain landmarks by two people
AIHub stroke follow-up imaging (Korea, dataset 71747)stroke follow-up imaging for thrombectomy decisions, same patient over time (modalities unverified)unverifiedunverifiedcontrolledAIHub terms; online safe zone, Korean nationals onlyaskneeds a request
the data needs a request or a data use agreement
found on AIHub; analysis only inside the online safe zone, no downloadlesion masks at each time point if present; two people place 15 landmarks on 10 pairs
AIHub stroke MR-CT synthetic pairs (Korea)synthetic brain MRI (DWI, SWI) and non-contrast CT pairs made by a model trained on real stroke pairs28,000 processed pairs of sliceslesion segmentations on CT and MRI (JSON polygons)both imagesaccountAI-Hub terms: 'only Korean nationals can apply for the data'; health data opened through the AI-Hub safe zone (analysis without download)askneeds an account
the download needs an account
Synthetic PNG pairs; real raw pairs stay in Korea's secure zone. Licence checked on 2026-10-10 at https://aihub.or.kr/aihubdata/data/view.do?aihubDataSe=data&dataSetSn=71964; derived test pairs shareable: no; commercial use: unclear. foreign company access not possible as stated; v1.1 opened 2026-06-30synthetic pairs are aligned by construction: apply a known warp
ALBERTs neonatal atlases (Gousias)Neonatal T1/T2 MR at term-equivalent age (15 preterm, 5 term); inter-subject20 neonates = 380 directed pairsManual 50-region segmentation by I. Gousiasboth imagesrequestRegistration form; academic non-commercial research only; no redistributionnon-commercialneeds a request
the data needs a request or a data use agreement
Link-out only, like Hammers n30r95 (same site). status change: Round 4 count: listed in 3 of the 38 round-4 surveys and benchmark papers (Fu 2020 review (PMB); Zou 2022 review (Front Oncol); Han 2022 GAN registration review).
Allen Human Brain Atlas (6 donors): MRI and sample coordinatesT1/T2/DTI MRI of six post-mortem donors with tissue-sample positions in native MRI and MNI space; between subjects to MNI6 donor-to-MNI pairsSample positions in native MRI (Allen) and three competing MNI coordinate sets (Allen, alleninf-style correction, 2025 refined set)both imagesdirectAllen Terms of Use (non-commercial by default)non-commercialtruth to make
the samples' MNI coordinates come from registrations (three published sets disagree), no independent truth; 6 donors.
Ding 2016 hemisphere and ds003590 already listed; this is the six-donor microarray set.Hundreds of sample points per donor in native MRI act as landmarks: map them with each method to MNI and compare against the refined 2025 set; two people check 30 points per donor
Allen Human Brain Reference Atlas (Ding 2016)One left hemisphere: 7T 200 µm MRI and DWI vs 641 Nissl + 287 parvalbumin sections, blockface1 specimen, ~900 sections to one MRIExpert annotations of ~860 structures on Nissl sectionsone imagedirectAllen Terms of Use: non-commercial by default, written permission for commercialnon-commercialtruth to make
no truth on both images; Two people place 10 landmarks (vessels, sulcal fundi, nucleus edges) on 10 Nissl section/MRI slice pairs, using blockface to find the slice
A search result claimed CC BY 4.0 since 2022; the Allen terms page (2022) says non-commercial. Treat as NC.Two people place 10 landmarks (vessels, sulcal fundi, nucleus edges) on 10 Nissl section/MRI slice pairs, using blockface to find the slice; non-commercial licence.
Amsterdam IMAGO glioma datasetadult glioma MRI (time points unverified)1,700 adultsunverifiedcontrolledCC BY-NC-ND 4.0 (article); access via EUCAIM catalogueaskneeds a request
the data needs a request or a data use agreement
Longitudinal content not checked.Check for follow-up scans first; if present, two raters place AFIDs away from the tumour on 10 pairs, SynthSeg labels on both as a weaker check.
AOMIC (ID1000, PIOP1, PIOP2)3T T1w, DWI and fMRI of one healthy subject in one session928 + 216 + 226 subjectsnone (FreeSurfer/fMRIPrep derivatives, automatic)directCC0opentruth to make
no truth on both images; warp T1w by a known field and register b0 (distortion-corrected) or mean EPI to it
Large, open, same-session. T1w vs b0/DWI is multimodal but EPI distortion and motion break exact alignment.warp T1w by a known field and register b0 (distortion-corrected) or mean EPI to it; exact truth only after field-map correction, else a known-warp test on top of a residual misalignment
APISNCCT and ADC MR of stroke, same patient60masks on ADC onlyone imageaccountCC BY-NC-SA 4.0non-commercialneeds an account
the download needs an account
Two people place 15 landmarks (ventricle horns, AFIDs points) on NCCT and ADC of 10 patients; SynthSeg labels on both (it reads CT and MR) as a weaker check.
B(RAIN)2 (DBS and iEEG resource, DABI)DBS patients: pre-op MRI (T1, T2, FGATIR, DWI), intra-op O-arm CT or fluoroscopy, pre- and post-op CT, ECoG/LFP; same patientunverified number of subjectsDBS lead and ECoG contact coordinates (Lead-DBS, manually adjusted) in native and MNI spaceone imageaccountnot readable: DABI pages are rendered client-side, no terms in the served HTMLunknownneeds an account
the download needs an account
DABI page says the repository is under review for modification. Access terms unverified. Licence checked on 2026-10-10 at https://dabi.loni.usc.edu/brain2/about; derived test pairs shareable: unclear; commercial use: unclear. DABI account needed to download; check its data use terms in a browserleads and ECoG strip contacts are bright on intra-op O-arm CT and post-op CT: contact centroids = landmarks between the two CTs; a stereotactic frame may also show on O-arm (unverified); limit: lead bending, brain shift
B-Q MINDED test-retest cross-scanner brain MRIHealthy volunteers, two 3T scanners, repeat sessionsUnverified countNoneone imagedirectnone stated (open access record, citation requested)unknowntruth to make
no truth on both images; Scan-rescan and cross-scanner pairs: near-identity, inverse and cycle consistency
Scan-rescan; truth fails Licence checked on 2026-10-10 at https://zenodo.org/api/records/6473268; derived test pairs shareable: unclear; commercial use: unclear. Zenodo open record with no licence field; description asks for two citations onlyScan-rescan and cross-scanner pairs: near-identity, inverse and cycle consistency; two raters place AFIDs on 10 cross-scanner pairs for point truth.
Baby Connectome Project (BCP)Longitudinal infant/child T1/T2/dMRI MR, birth to 5 yearsMany longitudinal pairsNone manual (the labelled part is released as BOBs)controlledNIMH Data Archive DUAaskneeds a request
the data needs a request or a data use agreement
Fails access; use BOBs instead.Use BOBs for labels; on BCP pairs, two raters place 20 landmarks (AFIDs adapted to infants) on 10 longitudinal pairs. NDA terms block redistribution.
BCBM-RadioGenomicsT1-post MRI of breast-cancer brain metastases at repeat Gamma Knife sessions, same patient months apart165 patients, 268 studies (about 100 repeat studies)Expert-reviewed masks of tumour, resection cavity and nearby structures on each studyboth imagesdirect: TCIA public IBM Aspera faspex package (no account), files one at a time; unstripped heads controlledCC BY 4.0 (skull-stripped images, segmentations, radiomics, clinical data); head images NIH Controlled Data Access Policyopenleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): open, but not among the six minor sets kept. It was proposed as: BCBM-RadioGenomics: 268 skull-stripped T1-post NIfTI of 165 patients, 60 of them with two or more Gamma Knife sessions, CC BY 4.0, 5.4 GB public Aspera package; the masks were created and reviewed by a neurosurgeon and a radiation oncologist. The collection's radiomics table shows they are almost all tumours and treatment targets, which change between sessions, while brainstem, chiasm or optic nerves are drawn on two sessions of only 8 patients, so the corpus would label the anatomy itself as for UCSDBrainMets, the same setting already ranked apart.
Single sequence; repeat sessions often treat new lesions, so label overlap across time is partial.
Booth lab ADHD MRI datasets (Lytle et al. 2020, 2021)T1 MP-RAGE, children with ADHD, between subjects47 children (LUMIR zero-shot)LUMIR-made labels (automatic)both imagesdirect: OpenNeuro, public S3 bucket s3://openneuro.org, file by file, no accountCC0 (dataset_description of ds002424 v1.2.0 and ds003500 v1.2.0 on OpenNeuro)openin the corpuspriority 3
The LUMIR companion paper (arXiv 2505.24160) pools 47 children with ADHD from two earlier studies (MP-RAGE at 3 T and 1.5 T) as a zero-shot set; the two Lytle et al. data papers are OpenNeuro ds002424 (79 children aged 8.6-12, 35 with ADHD, one T1w each, about 9 MB, 43 GB mostly fMRI) and ds003500 (38 subjects: 12 children with ADHD, 15 typical children, 11 adults), both CC0 on the open S3 bucket. LUMIR keeps its labels private (automated software, manual landmarks on a subset), so the corpus would label both images itself (FastSurfer, as for LUMIR) and rank it apart; it adds children to an adult-only LUMIR.
One of LUMIR 2025's zero-shot sets; OpenNeuro IDs not confirmed. Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (LUMIR 2025). Licence checked on 2026-10-10 at https://openneuro.org/datasets/ds002424; derived test pairs shareable: yes; commercial use: yes. other Booth-lab OpenNeuro sets with Lytle (ds003604, ds002236, ds001486) are also CC0; which ones LUMIR used is not confirmedAFIDs landmarks by two raters on 10 pairs
Brain Genomics Superstruct Project (GSP)T1 brain MRI of young adults, between subjectsabout 1,570 subjects in the open releasenone manualaccountGSP Open Access Data Use Terms (Harvard Dataverse, file access request): original and derived data may be redistributed under the same terms; no commercial clauseaskneeds an account
the download needs an account
Fails truth; training pool. Licence checked on 2026-10-10 at https://dataverse.harvard.edu/api/datasets/:persistentId/?persistentId=doi:10.7910/DVN/25833; derived test pairs shareable: yes; commercial use: unclear. terms version 2014-Apr-22; acknowledgment wording mandatory; commercial use not addressedSynthSeg labels on both, weaker check; AFIDs by two raters on 20 subjects
Brain metastases pre and post radiosurgery (ETZ Tilburg)T1c MRI at planning and follow-up after radiosurgerycount unverifiedmetastasis segmentations for a DL studyhiddenon site at ETZ after approvalaskleft out
its images or its truth are not released
Data cannot leave the hospital.Data cannot leave the hospital: unusable for a public benchmark.
Brain metastasis reirradiation longitudinal MRI benchmarkLongitudinal T1c MRI, same patient, SRS retreatment87 lesion pairsLesion segmentations at both time pointsboth imageshiddenInstitutionalaskleft out
its images or its truth are not released
Institutional cohort
Brain MRI full clinical protocol with and without intentional motion (OpenNeuro)six clinical brain sequences with and without deliberate head motion and with and without prospective correction, same session22 subjectshead motion from an external tracking device, shared as motion transforms; raw k-space tooboth imagesdirectunverified (OpenNeuro, usually CC0)unknowntruth to make
the tracker logs head motion within each scan for motion-correction work; a between-scan rigid truth would have to be made from it; rigid only.
Sci Data 2026. Tracker transforms are an independent rigid reference.verify tracker transforms with 10 landmarks per subject; known warps on motion-free scans for deformable tests
Brain Tumor Progression Challenge (BraTPRO, MICCAI 2024)longitudinal glioma MRI, consecutive visits of the same patient, RANO responsedevelopment set size unverified; hidden train and test setstumour segmentation masks in trainingaccountunverifiedunknownneeds an account
the download needs an account
classification challenge; the public development data may come from LUMIERE (already listed), unverifiedtwo raters place AFIDs away from the tumour on 10 visit pairs; tumour masks only as a weak check
Brain-Mets-Lung-MRI-Path-SegsPre-op MRI vs biopsy WSI103 patientsMRI segmentations onlyone imagedirectCC BY 4.0opentruth to make
no truth on both images; Biopsy tissue has no spatial link to the MRI: no usable truth.
No spatial correspondence; fails Licence checked on 2026-10-10 at https://www.cancerimagingarchive.net/collection/brain-mets-lung-mri-path-segs/; derived test pairs shareable: yes; commercial use: yes. TCIA table: NIfTI MR + segmentations CC BY 4.0, PublicBiopsy tissue has no spatial link to the MRI: no usable truth.
Brain-TR-GammaKnifeBrain metastases: Gamma Knife planning T1 MPRAGE + RT dose, follow-up outcome per lesion47 patients, 244 lesions; follow-up images not confirmed (unverified)Lesion RTSTRUCT on planning MRI; recurrence labels from follow-upone imagecontrolledNIH Controlled Data Access Policyaskneeds a request
the data needs a request or a data use agreement
Likely fails pairs (follow-up given as labels, not images). Images moved to dbGaP controlled access (phs004225, head images) in July 2025.Check first whether follow-up MRIs are released; if so, two raters place AFIDs away from lesions on 10 planning/follow-up pairs, SynthSeg labels on both as a weaker check.
Brain-Tumor-ProgressionGBM MRI at two time points (unverified)20 patientsTumour masks (unverified)both imagesaccountNIH Controlled Data Access Policyaskneeds an account
the download needs an account
Series co-registered to T1+C within each exam; small. Images moved to dbGaP controlled access (phs004225, head images) in July 2025. Licence checked on 2026-10-10 at https://www.cancerimagingarchive.net/collection/brain-tumor-progression/; derived test pairs shareable: unclear; commercial use: unclear. TCIA Limited access: images need an approved controlled-access request
BrainTRACE longitudinal brain MRI benchmarklongitudinal multi-sequence brain MRI studies with reports, same patient1,778 patients, 7,299 studies, about 29k co-registered volumesnone for registration (VQA benchmark)one imagehiddenunverifiedunknownleft out
its images or its truth are not released
Source and access not confirmed; volumes already co-registered.known warps on the co-registered volumes; lesion masks drawn on two visits for 20 patients
BrainWeb simulated brain MRIsimulated T1, T2 and PD MRI from digital brain phantoms (normal, MS, 20 anatomical models)20 models x contrasts, noise and bias levelsexact tissue labels shared by all contrasts of one phantomboth imagesdirectnone stated on the BrainWeb site (home, FAQ, about pages)unknowntruth to make
one simulated anatomy aligned across contrasts, so only a known warp (the known-warp track's kind); no licence stated; same data as 'BrainWeb Simulated Brain Database (normal and MS)'.
Aligned by construction, so a known warp gives exact multimodal truth. status change: Round 4 count: listed in 4 of the 38 round-4 surveys and benchmark papers (Boveiri 2020 review; Zou 2022 review (Front Oncol); Hammoudeh 2023 survey; Chen X 2021 review (Prog Biomed Eng)). Licence checked on 2026-10-10 at https://brainweb.bic.mni.mcgill.ca/brainweb/; derived test pairs shareable: unclear; commercial use: unclear. simulated data from the MNI; no licence or use terms publishedapply known smooth deformations between contrasts
BraTS 2021-2024 multi-parametric MRI (RSNA-ASNR-MICCAI)T1, T1c, T2, FLAIR of one patient in one exam; glioma (2021), plus meningioma, metastases, paediatric, Sub-Saharan Africa sets on Synapse (2023-2024)1,480 subjects on TCIA (2021); more on Synapse (counts unverified)tumour sub-region labels by experts, one label map valid on all four contrastsboth imagesdirectCC BY 4.0 (TCIA 2021 set); Synapse sets under Synapse terms (unverified)opentruth to make
contrasts pre-aligned with one label map for all, so only a known warp (the known-warp track's kind); 142 GB.
Contrasts are skull-stripped, resampled to 1 mm and co-registered by rigid registration to an atlas (SRI24 per BraTS papers, not stated on the TCIA page): aligned by a registration, not by the scanner. Tumour labels give a lesion check across contrasts. status change: usage counted from papers: same-subject contrasts used for synthetic-warp multimodal tests. status change: Round 4 count: listed in 8 of the 38 round-4 surveys and benchmark papers (Fu 2020 review (PMB); Boveiri 2020 review; Zou 2022 review (Front Oncol); Hammoudeh 2023 survey; Chen X 2021 review (Prog Biomed Eng); Han 2022 GAN registration review; Foundation models review 2026; Multimodal DIR benchmark 2026).warp T2 or FLAIR by a known smooth field and register T1/T1c to it: exact truth up to the original co-registration error; tumour labels warped with it give a lesion-overlap check
BraTS-PEDs (TCIA 2025)pediatric high-grade glioma, T1, T1CE, T2, FLAIR of the same session; between patients for inter-subject work457 patients (pre-treatment only)expert tumour segmentations on the co-registered sequencesboth imagesdirectunverified (TCIA collection licence)unknowntruth to make
pre-aligned sequences with one tumour map, so only a known warp; between patients the truth is the tumour alone; licence not stated.
Sequences are pre-aligned in BraTS style, so intra-patient pairs are aligned by construction.known smooth warps on the co-registered sequences for exact multi-contrast truth; tumour masks plus public segmenter labels for inter-subject
Buckner40 (FreeSurfer manual training set)T1 brain MRI, between subjects40 subjects (39 used in VoxelMorph)expert manual whole-brain labels (the FreeSurfer atlas training set)both imageshiddennot distributed as a public set (unverified)unknownleft out
its images or its truth are not released
Subset of the fMRIDC structural data. Widely cited test set but no public download found; papers from the MGH/MIT and Sydney groups had copies.use Mindboggle-101 or LPBA40 instead, which carry manual labels and are public
Burdenko-GBM-ProgressionLongitudinal GBM MRI with RT planning (unverified)180 patientsRT structures (unverified)accountNIH Controlled Data Access Policy (images); CC BY 4.0 (clinical CSV)askneeds an account
the download needs an account
MRIs in native space, RT objects on CT frame: contours on both images only for the follow-up subset. Images moved to dbGaP controlled access (phs004225, head images) in July 2025. Licence checked on 2026-10-10 at https://www.cancerimagingarchive.net/collection/burdenko-gbm-progression/; derived test pairs shareable: unclear; commercial use: unclear. TCIA Limited access for the DICOM imagesFor patients with follow-up MRI, two raters place AFIDs away from the tumour on 10 MRI pairs; for MR to planning CT, two people place 15 skull and ventricle landmarks; SynthSeg labels on both as a weaker check.
CAI2R simultaneous brain PET-MR raw dataset (Knoll 2017, NYU)10-minute brain FDG-PET and undersampled MPRAGE acquired simultaneously on a Biograph mMR, with coil maps, attenuation maps and scatter estimatessubject count not stated (likely one or few)PET and MR share scanner geometry (aligned by construction)requestnon-commercial academic use only; email agreement formnon-commercialneeds a request
the data needs a request or a data use agreement
raw data allow simulating PET/MR misalignment by reconstruction with shifted attenuation mapsapply a known rigid or smooth warp to the MR: exact multimodal truth
Calgary Preschool MRIChildren 2-8 years, T1/DTI/ASL/rsfMRI, 1-20 scans per child; same child over years431 datasets from 126 childrenNonedirectCC BY 4.0opentruth to make
no truth on both images; Two raters place AFIDs on 10 same-child pairs a year or more apart
Fails truth; good longitudinal growth pairs.Two raters place AFIDs on 10 same-child pairs a year or more apart; SynthSeg labels on both as a weaker check (growth changes shapes).
Calgary-Campinas CC-359T1 MR, 359 adults, 3 vendors, 1.5T and 3T; inter-subject and cross-scanner359 subjects; manual subsets belowManual brain masks (12), manual thalamus (60, 40 public), hypothalamus (38, hidden), STAPLE hippocampus and brain masks for all (consensus of automatic methods)both imagesdirect for the images and masks: https://sftp.conp.ca/users/calgary-campinas/ (CONP DataLad dataset, 'privacy: open'); the thalamus and hypothalamus benchmarks only through a Google formCC BY-ND 4.0 (download page: 'Our datasets are distributed under a Creative Commons Attribution-NoDerivatives 4.0 International Public License'; CONP DATS: CC BY-ND)askin the corpuspriority 3
In the corpus: the 347 subjects without a manual mask paired within each scanner, 17 test and 15 development pairs, FastSurfer labels (automatic), ranked apart; CC BY-ND, so built and scored locally, never shared, its starting alignments kept off the repository.
No-derivatives: can link and score, cannot redistribute modified images. status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Foundation models review 2026).
Cam-CAN (Cambridge ageing cohort MRI)3T T1w, T2w, DWI, MT, fMRI of one adult (18-88 y) in one session~650 imaged adults (unverified)nonerequestCamCAN Project Terms and Conditions (approved application): non-commercial ethical research for the proposed use only; data must stay on the institution's server, no copying to external storagenon-commercialneeds a request
the data needs a request or a data use agreement
Lifespan range makes atrophy a realistic difficulty axis. Licence checked on 2026-10-10 at https://camcan-archive.mrc-cbu.cam.ac.uk/dataaccess/; derived test pairs shareable: no; commercial use: no. applications without a specific hypothesis (e.g. 'AI and data mining') are rejected; derived data must be released open accesswarp T2w by a known field and register T1w of the same session: truth exact after rigid removal of inter-scan motion
CANDI ShareT1 MR of children, inter-subject103 to 150 (unverified)manual labelsboth imagesaccountCreative Commons Attribution (NITRC project licence field)openneeds an account
the download needs an account
NITRC account and click-through Licence checked on 2026-10-10 at https://www.nitrc.org/projects/candi_share; derived test pairs shareable: yes; commercial use: yes. NITRC login needed to download; CC BY version not given
CERMEP-IDB-MRXFDGMR, FDG-PET, CT of healthy subjects37automatic Hammers labels; modalities already co-registeredboth imagesrequestnone stated: 'bona fide researchers can request access via a short form'; data 'available from the corresponding author on reasonable request'askneeds a request
the data needs a request or a data use agreement
Licence checked on 2026-10-10 at https://pmc.ncbi.nlm.nih.gov/articles/PMC8446124; derived test pairs shareable: unclear; commercial use: unclear. ethics permission covers sharing on request; terms come with the approval
CFB-GBMGlioblastoma MRI at t0, ~4 and ~6 months, plus CT and RT data264 patientsTumour delineations/RT structures when availableone imagedirectCC BY 4.0opentruth to make
no truth on both images; Images are co-registered in pre-processing: use the raw images if released, then two raters place AFIDs away from the tumour on 10 pairs, SynthSeg labels on bot
Co-registered pre-processing; check raw Licence checked on 2026-10-10 at https://www.cancerimagingarchive.net/collection/cfb-gbm/; derived test pairs shareable: yes; commercial use: yes. TCIA table: NIfTI images and segmentations CC BY 4.0, PublicImages are co-registered in pre-processing: use the raw images if released, then two raters place AFIDs away from the tumour on 10 pairs, SynthSeg labels on both as a weaker check.
CFB-GBM v2.0multimodal MRI of glioblastoma at 3 time points, same patient264 patientstumour contours on all time points (97%, up from 35%), RANO labelsboth imagesdirect: TCIA public IBM Aspera faspex package, no account; files one at a time with ascp (as for BrainMetsMR)CC BY 4.0 (TCIA collection page: images, segmentations and every table)openleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): open, but not among the six minor sets kept. It was proposed as: TCIA CFB-GBM (version 4, 2026-09-11): 264 glioblastoma patients with T1Gd at t0 for all, at 4 months (t1) for 163 and 6 months (t2) for 118, every series rigidly co-registered to the t0 T1Gd and skull-stripped by the provider, CC BY 4.0, one public Aspera package of 208 GB read file by file. The truth people checked is the GTV at each time (v2.0: nnU-Net contours validated by five radiation oncologists, 97% complete), which changes with treatment, so as for UCSDBrainMets the corpus would label the anatomy itself and show the GTV; it would be the open glioma follow-up set next to BraTSReg (CC BY-ND, use ask).
status change: v2.0 (2026) completes tumour contours across t0-t2. Licence checked on 2026-10-10 at https://www.cancerimagingarchive.net/collection/cfb-gbm/; derived test pairs shareable: yes; commercial use: yes. arXiv 2608.17884 says v2.0 is on TCIA at the CFB-GBM collection; TCIA table lists NIfTI images and segmentations as CC BY 4.0two people place 15 landmarks on 20 pairs
CoBrA Lab high-resolution atlases (5 brains)3T T1/T2 MR at 0.3 mm (interpolated from 0.6 mm), 5 healthy adults; inter-subject5 brains = 20 directed pairsManual labels on the same 5 brains: hippocampal subfields (Winterburn), cerebellar lobules (Park), striatum/GP/thalamus (Tullo), hippocampal WM, amygdala, habenulaboth imagesdirectCC BY-NC-SA 4.0 (repo README)non-commercialto buildpriority 2
Manual hippocampal subfield, cerebellar and subcortical labels on 5 brains (10 pairs) at 0.3 mm; brain MR between subjects already covered with manual labels; NC-SA.
Densest manual subcortical/cerebellar labels on one small set.
COGEPISTIM (SEEG imaging, HIP)pre-implant MRI (T1, T2, FLAIR), post-implant MRI or CT with SEEG electrodes, post-explant MRI; same patient; epilepsy235 subjects with pre- and post-implant imaging; 81 also post-operativeSEEG contact coordinates in native and MNI152 space from the clinical localisationone imagecontrolledowned by CHU Grenoble Alpes; signed data transfer agreement requiredaskneeds a request
the data needs a request or a data use agreement
On the HIP collaborative space. Which image the native coordinates refer to, and whether any subject has both post-implant CT and MRI, is unverified.contacts are bright on post-implant CT and dark on post-implant MRI: contact centroids give exact same-session multimodal landmarks where both exist; pre vs post-implant MRI: two people place AFIDs; limits: brain shift after implantation, resection on post-op
Cornell negative BOLD and neurodegeneration multimodal PET/MRI (OpenNeuro ds006148)MRI (BOLD and structural) and PET of the same subjects266 subjectsnoneone imagerequestdata use agreement sent to the lab (README)askneeds a request
the data needs a request or a data use agreement
whether PET and MRI are simultaneous is unverifiedif simultaneous: aligned by construction, apply known warps; else FreeSurfer label transfer as a weak check
CoRR (Consortium for Reliability and Reproducibility)test-retest brain MR (T1, rs-fMRI, some DWI), many sites, same subject minutes to years apart1,629 subjects, 5,093 rs-fMRI scans from 18 sites (2014 paper); T1 retest count varies by site (unverified)noneaccountINDI terms: 'unrestricted usage of the datasets for non-commercial purposes' after NITRC registration; CoRR pages add only citation rulesnon-commercialneeds an account
the download needs an account
NITRC login with INDI group membership or COINS. Includes SLIM and CCNP baselines. Licence checked on 2026-10-10 at http://fcon_1000.projects.nitrc.org/; derived test pairs shareable: unclear; commercial use: no. per-site papers must be cited; no separate CoRR licencerescan pairs as near-identity, inverse and transitivity checks; AFIDs on a subset
CQ500 thick and thin CT with BHX hemorrhage boxes (PhysioNet)head CT, thick-slice vs thin-slice reconstructions of the same acquisitionCQ500 scans (count unverified); boxes on 23,409 imageshemorrhage bounding boxes, drawn on thick slices and extrapolated to thin slicesboth imagesdirectCC BY-SA 4.0 (BHX boxes)opentruth to make
thick and thin reconstructions of one acquisition, identity truth only (a sanity check).
same acquisition, so the true transform is identity up to slice geometry: a sanity check, not a deformable test; CQ500 has its own licenceapply known warps to the thin series; boxes give a weak check
CRCS-K stroke imaging repository (Korea)all stroke CT, MRI and angiography during hospital stay, 18 Korean centres20,792 patients, 225,159 seriesAI-derived lesion volumes, no correspondencesone imagecontrolledsteering committee approval; data stay in the platformaskneeds a request
the data needs a request or a data use agreement
Exports limited to aggregate outputs.Data stay in the platform and only aggregates leave: unusable for a public benchmark.
CRL fetal brain atlas (Gholipour 2017) and CRL-2025Fetal T2 MR atlas, 21-38 weeks, built from 81 (2017) and 160 (2025) fetusesAtlas only, no subject pairsAtlas tissue labels and 126-region parcellation (semi-automatic with manual refinement)directCC0 1.0 (IMAGINE fetal T2 atlas, 2017) and CC BY 4.0 (CRL2025 Fetal Dataset, T2W atlas, diffusion atlas) on Harvard Dataverseopentruth to make
193 fetal T2 reconstructions of 159 fetuses (31 scanned 2 or 3 times) are released in atlas space without labels; only the weekly atlases carry tissue labels
Fails: individual subjects not released; usable only as a template target. Licence checked on 2026-10-10 at https://dataverse.harvard.edu/api/datasets/:persistentId/?persistentId=doi:10.7910/DVN/A7YSH7; derived test pairs shareable: yes; commercial use: yes. Dataverse records have file access requests enabled; CRL lab page links themAtlas only, no subject images: a template target, no pair truth.
CT-MRI glioblastoma benchmark from Radiopaedia (IEEE DataPort)CT and MRI of glioblastoma, same patient, from Radiopaedia cases50 patientsnoneone imageaccountCC BY 4.0 on IEEE DataPort, but the cases are 'sourced from publicly available cases on Radiopaedia', which licenses cases CC BY-NC-SA 3.0non-commercialneeds an account
the download needs an account
Licence conflict; often partial series. Licence checked on 2026-10-10 at https://ieee-dataport.org/documents/ct-mri-glioblastoma-multimodal-benchmark-dataset; derived test pairs shareable: yes; commercial use: no. the upstream NC-SA terms cannot be relaxed by the repackager; IEEE DataPort login needed; 2D JPEG imagestwo people mark 10 landmarks per pair
CuRIOUS 2022 / RESECT-SEGPre-op MRI vs iUS before/during/after resection, same patient23 RESECT cases (training); hidden test casesRESECT-SEG: tumour in pre-resection iUS, cavity in during/after iUS (Behboodi et al.)one imagedirectchallenge page: images 'subject to copyright of the authors, and cannot be used elsewhere without permission'asktruth to make
no truth on both images; Draw the tumour on pre-op MRI to pair with the RESECT-SEG tumour on pre-resection iUS (contours on both)
status change for RESECT: new US segmentations (osf.io/6y4db); iUS-iUS pairs gain labels Licence checked on 2026-10-10 at https://curious2022.grand-challenge.org/data/; derived test pairs shareable: unclear; commercial use: unclear. original RESECT release reported as CC BY 4.0 (not re-checked); the challenge segmentations need permissionDraw the tumour on pre-op MRI to pair with the RESECT-SEG tumour on pre-resection iUS (contours on both); cavity labels on during/after iUS have no pre-op counterpart. CuRIOUS test data needs the authors' permission.
Cyprus brain metastases longitudinal (PROTEAS)Longitudinal MRI and CT, same patient40 patients, 185 studiesManual segmentations of 65 metastases (enhancing, edema, necrosis); timepoints unverifieddirect: Zenodo record 23171182 (version 4 of concept 15499111; the paper cites version 17253793), one zip a patient, readable by rangesCC BY 4.0 on the Zenodo deposit (the Sci Data article itself is CC BY-NC-ND)openleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): open, but not among the six minor sets kept. It was proposed as: Zenodo version 4 of the PROTEAS data (CC BY 4.0, 15.1 GB, one zip a patient): 40 patients, a baseline and at least one follow-up MRI (scheduled at 6 weeks to 12 months) (T1, T1c, T2, FLAIR) resampled by the provider into the BraTS/SRI24 frame (240x240x155, 1 mm), with metastasis masks at every visit made by networks or Imalytics and corrected by a neuroradiologist (Sci Data 2025). The lesions change under radiosurgery, so the masks can only be other labels: the ranked truth would be SynthSeg labels on both dates (automatic) as for UCSDBrainMets, whose atlas-aligned pairs leave a Dice of 0.80 as delivered. The same files would also give a brain-metastasis segmentation set (40 patients) and MR-to-CT synthesis pairs (45 planning CTs aligned to the T1 by the provider's non-rigid NiftyReg registration).
Published 2026-10-05Two raters place AFIDs away from lesions on 10 longitudinal pairs; metastases segmented at both time points give a lesion-centroid check (lesions shrink); SynthSeg labels on both as a weaker check.
Dallas Lifespan Brain Study (OpenNeuro ds004856)T1, T2, DWI, fMRI plus amyloid and tau PET, adults 20-90, same subject over 3 waves ~4-5 years apart464 subjects; 314 with MRI at 2+ waves (MRI per wave 464/300/193, from participants.tsv)nonedirectCC0opentruth to make
no truth on both images; two raters place the 32 AFIDs fiducials (published protocol) on both scans of 20 pairs (wave 1 vs wave 3)
Wide age range gives real atrophy change; same-subject MR-PET pairs too.two raters place the 32 AFIDs fiducials (published protocol) on both scans of 20 pairs (wave 1 vs wave 3); SynthSeg labels on both as a weaker check
Day2dayT1, high-resolution hippocampal T2, DTI, rs-fMRI; 8 adults scanned 40-50 times over 6-8 months on one scanner8 subjects (6 complete), ~300 T1 sessions (11-50 per subject)nonerequestno licence: dataset 'made available for public scientific use' (screening note: data sent by the corresponding author)non-commercialneeds a request
the data needs a request or a data use agreement
One scanner and protocol: a drift-free dense rescan series. Licence checked on 2026-10-10 at https://pmc.ncbi.nlm.nih.gov/articles/PMC5571657; derived test pairs shareable: unclear; commercial use: unclear. scientific-use wording only; no terms on redistribution or commercial usenear-identity and transitivity over 40-50 scans; hippocampal landmarks on the high-resolution T2
devCCNP (Chinese Color Nest Project)T1, rs-fMRI, DTI of children and adolescents 6-18, same subject at up to 3 waves479 participants, 864 visits (first release)nonerequestData Use Agreement on CCNP (DUA-CCNP), scientific research onlynon-commercialneeds a request
the data needs a request or a data use agreement
Baseline CKG sample is also in CoRR (IPCAS 7).two raters place the 32 AFIDs fiducials (published protocol) on both scans of 20 pairs; SynthSeg labels on both
dHCP neonatal and fetal releasesNeonatal T1/T2 MR (887 datasets from 783 infants, some scanned twice) and fetal MR; same subject (preterm at birth vs term) and inter-subject~100 longitudinal neonatal pairs (unverified count); many inter-subjectDrawEM multi-atlas labels (automatic)both imagesaccountdHCP data sharing agreement; no commercial purpose (incl. paid services)non-commercialneeds an account
the download needs an account
Fails expert truth (automatic labels). Commercial-service clause rules out a paid benchmark service.
Duke longitudinal glioblastoma MRI with landmark pairsPre-op vs recurrence multi-sequence MRI, same patient61 patients, 503 sessions; landmarks on pre-op and recurrence scansVessel-bifurcation landmark pairs (near/far from tumour), Duke group (Criscuolo, Calabrese)both imagesrequestrestricted record, no licence shownaskneeds a request
the data needs a request or a data use agreement
Same group as Lung/Liver/Abdominal-DIR-QA; images MNI co-registered and skull-stripped Licence checked on 2026-10-10 at https://zenodo.org/api/records/20548528; derived test pairs shareable: unclear; commercial use: unclear. Zenodo restricted access; request needed, terms not public
DWI Traveling Human Phantom Study (OpenNeuro ds000206)diffusion MR of 6 travelling subjects at 8 sites (CCF, Dartmouth, Iowa, JHU, MGH, UCI, UMN, UW)6 subjects, one or more sessions per sitenonedirectCC0opentruth to make
no truth on both images; cross-site near-identity
Anatomical scans present unverified.cross-site near-identity; FA-based landmarks (weak)
EADC-ADNI HarP hippocampus labelsADNI T1 MR at 1.5T and 3T; inter-subject and same subject across field strength135 ADNI scans; 10 subjects with 1.5T and 3T both tracedManual hippocampus labels, Harmonized Protocol, certified tracersboth imagescontrolledNot stated; via LONI IDA login (ADNI terms)askneeds a request
the data needs a request or a data use agreement
Labels done on MINC; NIfTI conversion may not match exactly (page warning).
eddymotionPET benchmark (simulated PET head motion)simulated dynamic PET for estimating motion between frames5 subjectssimulated inter-frame motion (whether the motion parameters are shared is unverified)directCC0opentruth to make
five simulated dynamic brain PET series; the simulated motion is not shipped, only FSL, AFNI and petprep_hmc estimates
README is one line; content and truth files not checked.if motion parameters are included: rigid frame-to-frame truth; else add known rigid motions to frames and score recovery
EPISURGPre-op T1 MR vs post-op T1 MR after epilepsy resection; same subject269 pre/post pairs (430 post-op)Resection cavity masks on post-op by 3 raters (133+34+33 subjects, overlapping)one imagedirectCC BY-NC-SA 4.0 (figshare API)non-commercialtruth to make
no truth on both images; Two raters place AFIDs away from the cavity on 10 pre/post pairs
Missing-tissue registration; cavity mask only on post-op, so it scores exclusion handling, not alignment. Defaced.Two raters place AFIDs away from the cavity on 10 pre/post pairs; SynthSeg labels on both, outside the cavity, as a weaker check.
Esteban diffusion MRI phantoms (RegSeg)simulated diffusion MRI with known susceptibility distortion vs T1w structurecount unverifiedknown distortion fields and surfaces from simulationboth imagesdirectCC BY 4.0openleft out
the figshare record holds only the paper's supplemental PDF; the phantoms are simulated.
Digital phantom; exact DWI-to-T1 truth.
Ex vivo medial temporal lobe MRI atlas with serial histology (Penn, OpenNeuro ds004767 and ds003052)9.4 T ex vivo MRI of temporal lobe specimens, between specimens, plus tau maps from serial histology55 specimens (25 in the 2021 release ds003052)MTL subregion segmentations per specimen and a consensus atlas labelling from 17 specimens (Penn group)both imagesdirect: OpenNeuro S3 bucket (s3://openneuro.org/ds004767), 65 to 180 MB a segmentation-space T2wCC0 (dataset_description.json of both releases)openin the corpuspriority 2
ds004767 (CC0) holds 9.4 T T2w scans of 55 temporal lobe specimens (34 CNDR, 21 HNL) with, for each, the MTL segmentation in derivatives/manual_segmentation, which the earlier release calls semi-manual; one sample label map (0.2 mm, 560 x 336 x 238) holds three labels. ds003052, the 2021 release of 25 specimens, shares 22 specimen ids with it (its three NDRI specimens are only there), so build ds004767 and at most add those three. Ex vivo specimens between donors are a setting the corpus lacks; whether the start leaves room (Dice under 0.8 after the affine) is to be measured.
per-specimen labels may be automatic propagations for some specimens; check the manifestlabel overlap between specimens; two people mark hippocampal head, body and tail landmarks on 10 specimens
Ex vivo MRI and neuropathology in healthy and VCID brains (DANDI 001437)7T MRI and 3T diffusion MRI of ex vivo brain samples vs neuropathology images, same specimensubject count unverified (asset summary lists 2)nonedirectCC BY 4.0opentruth to make
no truth on both images; two people mark 15 vessel and sulcal landmarks on MRI and on matching histology sections of 10 samples
found through the DANDI APItwo people mark 15 vessel and sulcal landmarks on MRI and on matching histology sections of 10 samples
fastMRI and fastMRI+ (brain, knee, prostate, breast raw MRI)raw k-space MRI: brain T1/T1post/T2/FLAIR, knee PD and PD fat-sat, prostate T2 + DWI per exam, breast DCE (GRASP)312 prostate exams with T2 and DWI; ~7,000 brain and ~1,400 knee raw scans (contrasts per patient not guaranteed)fastMRI+: bounding boxes of pathology by one radiologist (knee and ~1,000 brain)one imagerequestfastMRI data sharing agreement, academic use (per fastMRI+ README)non-commercialleft out
fastMRI data sharing agreement (read 2026-10-16): no distribution, publication or transfer; every user must register individually
Raw brain/knee files are one contrast each; same-patient multi-contrast pairing is not guaranteed (unverified). Prostate T2+DWI and breast DCE frames come from one exam: best use here. DWI has EPI distortion. status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Foundation models review 2026).prostate: warp T2 by a known field and register DWI/ADC to it (exact only up to EPI distortion and patient motion); breast DCE: warp one frame of the same radial acquisition
FBIRN / BIRN PD-weighted brain scans (SynthMorph test set)PD-weighted 1.5 T brain MRI, between subjects8 subjectsmanual brain segmentationsboth imageshiddenunknownunknownleft out
its images or its truth are not released
Origin of the files not stated beyond the BIRN reference; probably not distributed.
FeTA (KISPI training set, FeTA 2.x)Fetal T2 MR, super-resolution reconstructed, 21-36 weeks; inter-subject80 training subjects; inter-subject pairs, best within the same gestational weekManual labels of 7 tissues (organisers); FeTA 2024 adds clinician landmarks for 5 biometric measures (endpoints)both imagesaccountCustom: research and education only; KISPI data non-commercial (FeTA 2024 paper)non-commercialneeds an account
the download needs an account
Vienna training data is challenge-only (DTA). Whether the biometry landmarks are on Synapse outside the challenge: unverified. Synapse login plus terms. status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Azad 2023 ViT review).
Fetal brain ultrasound video and MRI pairs (VIFBA)fetal brain ultrasound videos and MRI of the same fetus857 cases, 3,196 videosMRI-derived ventricle volumes; no spatial correspondenceone imagehiddennot statedunknownleft out
its images or its truth are not released
Likely private; no release statement seen.ventricle and midline landmarks placed by two people on US planes and MRI
Fetal cerebellum landmark benchmark (APGA, JBHI 2025)Thin-slice 3D fetal MR, cerebellum landmarks; inter-subjectunverifiedCerebellum landmarks (authors' clinicians)both imageshiddenunverifiedunknownleft out
its images or its truth are not released
Paper says it introduces a benchmark; no public download found. Ask the authors.
Fetal spina bifida spatio-temporal atlas (Fidon 2022)Fetal T2 MR atlases per gestational week (21-34), spina bifida and control atlases; atlas to atlas or atlas to subjectAtlas volumes only (one per week and group)Tissue segmentations and anatomical landmarks (lmks.nii.gz) on the atlases, made by the authorsboth imagesdirect: Zenodo (also on Synapse syn25887675, which needs an account), no account on Zenodoatlas CC0 1.0, code BSD-3-Clause (Zenodo README and LICENSE in the zip)openleft out
The 13.6 MB zip holds 15 atlas volumes, one per gestational week and group (spina bifida not operated GA21-25, operated GA25-34), each with a super-resolution T2 (srr), a mask, a tissue parcellation and landmarks (lmks.nii.gz). Pairs would be between population averages of consecutive weeks, not between subjects; CC0 for the atlas, BSD-3-Clause for the code (Zenodo README).
No individual subjects; landmarks per atlas volume, protocol in the paper. Pairs across weeks are weak; better as targets for FeTA subjects. Also on Synapse syn25887675.
FieldTrip iEEG subject UCI29 (Stolk 2018)pre-implant T1, post-implant MRI and post-implant CT with ECoG grids and depth electrodes, same patient1 subject (3 images)electrode positions in ACPC (before and after brain-shift correction), MNI and fsaverage, by the authorsone imagedirectCC BY-SA 4.0opentruth to make
no truth on both images; apply a known rigid or smooth warp to the fused CT for exact truth
CT and post-implant MRI are released already fused to the ACPC MRI ('_f' files). One subject only; also in the Donders repository.apply a known rigid or smooth warp to the fused CT for exact truth; contacts seen on both CT and post-implant MRI check the fusion
FS40 (Buckner40) manually labelled cortical surfacesT1 MR of 40 subjects (young, middle-aged, elderly, Alzheimer's), between subjects40 subjectsmanually labelled cortical surfaces (Desikan-Killiany protocol) at MGHboth imageshiddenraw scans from an open Buckner-lab set; manual labels not distributed (unverified)unknownleft out
its images or its truth are not released
Raw T1 of these subjects appear to ship with FreeSurfer tutorial data (unverified); the manual labels do not.labels from FreeSurfer or another public segmenter on both images, a weaker check; or ask MGH
GammaKnife-HippocampalGamma Knife planning CT + T1 MRI (other sequences and follow-up MRI when available), same patient390 patients; 872 MR-to-CT REG objectsRigid MR-to-CT registrations made in MIM (DICOM REG); hippocampus contours by several institutional observers; planning structurescontrolledNIH Controlled Data Access Policyaskneeds a request
the data needs a request or a data use agreement
Rigid only; reference registration is clinical. Follow-up MR count not verified. Images moved to dbGaP controlled access (phs004225, head images) in July 2025.If the stereotactic frame is in the images, its fiducial rods on CT and MRI give marker truth; otherwise two people place 15 skull and ventricle landmarks on 10 MR-CT pairs. The MIM registration is a clinical reference, not truth.
GLIS-RTPost-op glioma: planning CT + MRI (3D T1c, T2 FLAIR) of the same patient (CT/MR)230 CT/MR pairsClinical CT-MR registrations (DICOM REG); GTV/CTV and anatomical barrier and organ contourscontrolledNIH Controlled Data Access Policyaskneeds a request
the data needs a request or a data use agreement
REG is the clinical registration: a reference, not independent truth. Images moved to dbGaP controlled access (phs004225, head images) in July 2025.Two people place 15 landmarks (skull, ventricles, vessels) on CT and MRI of 10 patients to score the rigid fit; SynthSeg labels on both as a weaker check. The clinical REG is a reference, not truth.
HABS (Harvard Aging Brain Study)T1 brain MRI of older adults, longitudinal and between subjectsabout 280 baseline subjects, follow-ups (unverified)none manualrequestpublic data release v3 on Synapse 'upon request and agreement to our data use agreement (DUA)'; DUA text not on the HABS siteaskneeds a request
the data needs a request or a data use agreement
Fails truth. Licence checked on 2026-10-10 at https://habs.mgh.harvard.edu/researchers/; derived test pairs shareable: unclear; commercial use: unclear. apply via www.synapse.org/habsrepeat visits as near-identity checks; AFIDs on 10 longitudinal pairs
Hammers n30r95T1 MR, inter-subject30 subjects95 manually drawn regionsboth imagesaccountacademic non-commercial research, no redistributionnon-commercialneeds an account
the download needs an account
registration form
HCP Lifespan (HCP-Aging and HCP-Development)T1w/T2w brain MRI across ages 5-100, between subjectsabout 1,300 + 1,300 subjects (unverified)none manual (SAMSEG/FreeSurfer in papers)both imagescontrolledNIMH Data Archive data use certificationaskneeds a request
the data needs a request or a data use agreement
Fails truth and open access. Same-session T1w and T2w give aligned pairs.known warp on the same-session T2w gives exact T1-T2 truth; AFIDs on 20 subjects
HCP Young Adult S1200 (full release)T1w/T2w and diffusion MRI, between subjects or T1-to-dMRI within subjectabout 1,100 subjects (1,076 used by uniGradICON)FreeSurfer labels (automatic)both imagesaccountWU-Minn HCP Open Access Data Use Terms (ConnectomeDB account, click-through): original and derived data may be redistributed under the same terms; no commercial clauseaskneeds an account
the download needs an account
Broader than the 'HCP Young Adult test-retest' entry already listed. Same-session T1w, T2w and dMRI are aligned by acquisition and preprocessing, so a known warp gives multimodal truth. Licence checked on 2026-10-10 at https://www.humanconnectome.org/study/hcp-young-adult/document/wu-minn-hcp-consortium-open-access-data-use-terms; derived test pairs shareable: yes; commercial use: unclear. everyone using shared copies must register and accept the terms; IRB approval may be needed; commercial use not addressed, confirm before a paid serviceapply a known smooth warp to the T2w or b0 of the same session: exact multimodal truth; AFIDs by two raters on 20 subjects for inter-subject points
HCP Young Adult test-retestT1/T2 0.7 mm MR, 45 subjects scanned twice months apart45 rescan pairs; 1,100 subjects for inter-subjectFreeSurfer (automatic)both imagesaccountWU-Minn HCP Open Access Data Use Terms (ConnectomeDB account, click-through): original and derived data may be redistributed under the same terms; no commercial clauseaskneeds an account
the download needs an account
Fails expert truth; rescan pairs give a near-identity check. status change: Round 4 count: listed in 4 of the 38 round-4 surveys and benchmark papers (Foundation models review 2026; uniGradICON 2024; multiGradICON 2024; Multimodal DIR benchmark 2026). Licence checked on 2026-10-10 at https://www.humanconnectome.org/study/hcp-young-adult/document/wu-minn-hcp-consortium-open-access-data-use-terms; derived test pairs shareable: yes; commercial use: unclear. same terms as S1200; team members must each accept the terms
HEAR-MRI head fiducialsstructural head MRI from 9 EEG-MRI cohorts, between subjects342 scansnasion and helix-tragus fiducials per scan, intra-rater ICC 0.93both imagesdirect: Zenodo (labels) and OpenNeuro (images), no accountCC BY 4.0 (Zenodo) for the labels; images under each OpenNeuro dataset's licenceopenleft out
Version 1.0.1 (CC BY 4.0) releases only labels and benchmark code (1.9 MB): the nasion and the two helix-tragus junctions placed by hand on 342 T1 MRIs of nine OpenNeuro EEG-MRI cohorts, the images staying on OpenNeuro, and the nasion void in 99 of 342 heads because of defacing. Three points a head, often two, cannot score a deformable registration; AFIDs already gives 32 fiducials a brain.
Only 3 points per head; useful for head surface or EEG-MRI alignment, weak for brain.
hMRI toolbox MPM example dataset3T multi-parameter mapping: multi-echo PDw, MTw, T1w FLASH plus B1/B0/receive maps, one subject1 subject; 3 multi-echo volumes (6-8 echoes each)none; deliberate head rotation logged between PDw and MTwdirectCC BY 4.0opentruth to make
no truth on both images; warp PDw (or one echo) by a known field and register T1w/MTw echoes to it
Echoes of one FLASH are aligned by construction. Subject rotated about z before MTw and returned before T1w: a real-motion multi-contrast pair, magnitude not measured.warp PDw (or one echo) by a known field and register T1w/MTw echoes to it; the rotated MTw is a real-motion case scored by rigid consistency (PDw->MTw->T1w loop)
HOA-2 subcortical parcellations of HCP subjects (Harvard-Oxford Atlas 2.0)T1 brain MRI of HCP Young Adults, between subjects100 subjects in v2.0.0 (50 in v1), any pairmanual anatomically curated subcortical labels (28 regions; cerebellum added in v2.1.0) by the MGH Center for Morphometric Analysisboth imagesaccountlabels: 3D Slicer license part B; images: HCP Open Access Data Use Termsaskneeds an account
the download needs an account
Manual expert labels on public HCP1200 subjects: one of the few large manual inter-subject brain truths. Labels on Zenodo (doi 10.5281/zenodo.7948708); images must be pulled from HCP. Licence checked on 2026-10-10 at https://github.com/HOA-2/SubcorticalParcellations; derived test pairs shareable: unclear; commercial use: unclear. README: MRB files with HCP data not posted because of HCP data use agreements, by request
Human es-fMRI resource (OpenNeuro ds002799)pre-op T1 MRI and post-implant T1 MRI with depth and grid electrodes, same patient (plus es-fMRI)26 patients with pre-op and post-implant T1stimulation electrode coordinates in MNI space onlyone imagedirectCC0opentruth to make
no truth on both images; two people place AFIDs (32 points) on pre-op and post-implant T1 of 10 pairs
Electrodes are artefacts on post-op T1 only; no native-space coordinates; brain shift from implantation is real.two people place AFIDs (32 points) on pre-op and post-implant T1 of 10 pairs; electrode artefacts give no cross-image truth
IBIS (Infant Brain Imaging Study)infant T1/T2 MR, longitudinal, between or within subjectsunverifiednone for registrationcontrolledNIMH Data Archive terms (unverified)askneeds a request
the data needs a request or a data use agreement
Listed by Fu 2020 and Hammoudeh 2023; used by Quicksilver-type papers. Project page did not answer (404). Round 4 count: listed in 3 of the 38 round-4 surveys and benchmark papers (Fu 2020 review (PMB); Hammoudeh 2023 survey; Chen X 2021 review (Prog Biomed Eng)).infant segmentations (e.g. iBEAT) on both visits, weak; landmarks by two raters on 10 pairs
IBSR v1.0 (20 normal subjects)raw T1 MR of healthy adults, between subjects20 subjects (Ou 2014 used 10, 90 pairs)manual brain masks and tissue segmentations by the MGH Center for Morphometric Analysisboth imagesaccountunverified (the listed IBSR18 entry says non-commercial only)non-commercialneeds an account
the download needs an account
Not the IBSR18 set of Klein 2009 (listed). Truth is brain mask and tissue classes only, a weak check for deformable registration.labels from a public segmenter (SynthSeg, FreeSurfer) on both images, a weaker check; or 15 landmarks per scan by two people
IDEAS (Imaging Database for Epilepsy And Surgery)Pre-op T1/FLAIR MR of 442 epilepsy surgery patients + 100 controlsNo pairs (post-op images not released)Resection masks in pre-op space, drawn from post-op imaging; FreeSurfer parcellations manually inspectedone imagedirectCC0 (OpenNeuro); masks on figshareopentruth to make
no truth on both images; No post-op image is released: no pair truth
Fails: no post-op image. Only inter-subject with automatic parcellations.No post-op image is released: no pair truth; inter-subject only, with AFIDs by two raters on 10 pairs.
Individual Brain Charting (IBC, OpenNeuro ds002685)12 adults with many sessions of high-resolution fMRI; T1, T2, FLAIR, DWI13 subjects, ~45 sessions each; anatomical scan per session unverifiednonedirectCC0opentruth to make
no truth on both images; same-session T1/T2/FLAIR give multimodal pairs under a known warp
README says T2 and FLAIR were screening scans; anatomy may be sparse per session.same-session T1/T2/FLAIR give multimodal pairs under a known warp
Infant FreeSurfer manual set (de Macedo Rodrigues 2015)Infant T1 MR, 0-2 years; inter-subjectunverified (about two dozen subjects)FreeSurfer-compliant manual segmentationsboth imagesrequestnone stated: paper only plans to release an atlas from the training data as part of FreeSurferunknownneeds a request
the data needs a request or a data use agreement
Availability not checked; ask MGH (Zöllei). Licence checked on 2026-10-10 at https://pmc.ncbi.nlm.nih.gov/articles/PMC4332305; derived test pairs shareable: unclear; commercial use: unclear. no public download of the manual label set found
Insight MS longitudinal MRIT1, FLAIR and T2 of MS patients at 2 or more time points172 patients, 496 scanslesion and new-lesion labels with three-stage expert reviewboth imagesrequestrestricted record, no licence shownaskneeds a request
the data needs a request or a data use agreement
Lesion masks track change, not anatomy. Licence checked on 2026-10-10 at https://zenodo.org/api/records/8146425; derived test pairs shareable: unclear; commercial use: unclear. Zenodo restricted access; request needed, terms not publictwo people place 15 anatomical landmarks on 10 pairs
ISBI 2015 longitudinal MSMR at several visits, multiple sclerosis, same patient5 training subjectslesion masks by 2 ratersboth imagesaccountterms not readable: IACL page says data come from the Smart Stats website after creating an account; smart-stats-tools.org does not respondunknownneeds an account
the download needs an account
test masks hidden Licence checked on 2026-10-10 at https://iacl.ece.jhu.edu/index.php?title=MSChallenge/data; derived test pairs shareable: unclear; commercial use: unclear. no licence or use terms on the IACL challenge pages
iSeg-2017 / iSeg-20196-month infant T1/T2 MR (MAP study, multi-site in 2019); inter-subject10 labelled training subjects (2017) plus 2019 multi-site validationWM/GM/CSF labels, automatic then manually corrected by expertsboth imagesrequestchallenge rules: data only for preparing an iSeg entry; no other scientific use, no training of other algorithms; no redistribution of data or derived dataaskneeds a request
the data needs a request or a data use agreement
Only 3 tissue classes; T1 and T2 already aligned. Licence checked on 2026-10-10 at https://iseg2017.web.unc.edu/rules/; derived test pairs shareable: no; commercial use: no. iSeg-2019 site gives no separate terms; registration required
ISLES 2015multi-sequence stroke MR, same patient (sequences)~64 casesmanual lesion masksone imageaccountnone stated on the challenge page; data distributed through the VSD/SMIR platform (smir.ch now 403/404), whose dataset terms could not be readunknownneeds an account
the download needs an account
Listed by Fu 2020 and Boveiri 2020 (rigid multi-sequence registration). Fails truth: lesion masks are not registration truth. Round 4 count: listed in 2 of the 38 round-4 surveys and benchmark papers (Fu 2020 review (PMB); Boveiri 2020 review). Licence checked on 2026-10-10 at http://www.isles-challenge.org/ISLES2015/; derived test pairs shareable: unclear; commercial use: unclear. challenge rules only cover participationsequences of one session are near-aligned: apply a known warp to one sequence for exact multimodal truth
ISLES'24CT and MR of acute stroke, same patient150lesion labels on CT and on MR are different structures; MR supplied in CT spaceone imagedirectCC BY-NC-SA 4.0non-commercialtruth to make
no truth on both images; MR is supplied resampled to CT, so the pairs are already aligned: no usable truth unless native MR is released.
MR is supplied resampled to CT, so the pairs are already aligned: no usable truth unless native MR is released.
IXI (TransMorph preprocessing)T1 MR to an atlas, inter-subject576 subjectsFreeSurfer labels (automatic)both imagesdirect: one 1.44 GB file on Google Drive (public link, no account; Drive's large-file warning page only)CC BY-SA 3.0 (TransMorph_on_IXI.md: 'Our preprocessed IXI dataset is made available under the Creative Commons Attribution-ShareAlike 3.0 Unported License'; IXI itself CC BY-SA 3.0)openin the corpuspriority 2
Built: IXI_TransMorph (track ixi-atlas-automatic).
Truth made by a program. status change: usage counted from papers: the atlas-to-subject split (403/58/115) is the de facto second benchmark after Learn2Reg OASIS; multi-contrast users take the raw IXI T1/T2/PD instead. status change: Round 4 count: listed in 11 of the 38 round-4 surveys and benchmark papers (Fu 2020 review (PMB); Boveiri 2020 review; Zou 2022 review (Front Oncol); Hammoudeh 2023 survey; Chen X 2021 review (Prog Biomed Eng); Chen 2025 survey (MedIA); Azad 2023 ViT review; Han 2022 GAN registration review; uniGradICON 2024; multiGradICON 2024; rethink-reg 2024).labels are automatic: two raters place the 32 AFIDs on the atlas and 20 test subjects; same-session T2/PD give exact multimodal truth under a known warp
IXI multi-contrast (T1, T2/PD dual echo, MRA, DTI)T1, T2, PD, MRA and DTI of one healthy subject in one session; 3 London sites (Philips 3T/1.5T, GE 1.5T)~600 subjects; T2 and PD come from one dual-echo acquisitionnone; T2/PD aligned by acquisitiondirectCC BY-SA 3.0opentruth to make
no truth on both images; warp the PD (or T1) by a known smooth field and register T2 to it: exact T2-PD truth
status change: listed only as 'IXI (TransMorph preprocessing)'; the multi-contrast intra-subject use was not. T2 and PD share the echo train so they are aligned exactly; T1 and MRA are separate scans and move a little.warp the PD (or T1) by a known smooth field and register T2 to it: exact T2-PD truth; for T1-T2 first check residual motion, rigidly align, then warp
J-ADNI (Japanese ADNI)longitudinal MRI and PET of a Japanese Alzheimer cohortparticipants and visits unverifiednoneone imagecontrolledNBDC controlled access (unverified)askneeds a request
the data needs a request or a data use agreement
Reported to be shared through the NBDC Human Database; record not found.labels from a public segmenter on both visits (weaker)
Kirby21 / Multi-Modal MRI Reproducibility Resource21 healthy adults scanned twice in one day (MPRAGE, FLAIR, DTI, qT1/qT2, etc.)21 rescan pairsNoneaccountBIRN Data Repository Data Use Agreement (NITRC licence field): free with e-mail and acceptance; originals redistributable under the same terms, derived data under the deriver's terms; no commercial clauseaskneeds an account
the download needs an account
Fails truth. Licence checked on 2026-10-10 at https://www.nitrc.org/projects/multimodal/; derived test pairs shareable: yes; commercial use: unclear. acknowledgement and citation required; no purpose or commercial restriction in the DUA textSame-day rescans: near-identity, inverse and cycle consistency; same-session MPRAGE, FLAIR and qT1 under a known warp give near-exact multi-contrast truth; AFIDs by two raters on 10 rescan pairs.
Lead-Tutor (DBS electrode localisation)pre-op MRI (T1, T2, some FLAIR) and post-op CT with bilateral DBS leads, same patient10 patients (20 leads)lead head, tail and orientation in MNI space, average of two experts; 10 trained non-experts give the spreadone imagedirectnone stated (OSF project public, no licence set; title says 'Open Access Dataset')unknowntruth to make
no truth on both images; leads show on CT only
Expert lead positions came through the experts' own MR-CT and MNI registrations, so they check the whole chain, not MR-CT alone. Data on OSF (paper link is view-only). Licence checked on 2026-10-10 at https://api.osf.io/v2/nodes/jxecg/; derived test pairs shareable: unclear; commercial use: unclear. OSF files LeadTutor.zip + description PDF, neither states termsleads show on CT only; for MR-CT truth two people place ~12 bony and vascular landmarks (AC, PC, foramina, skull points) on MR and CT; lead contacts by PaCER then a person give the CT side of the chain; limit: pneumocephalus brain shift
Ljubljana longitudinal MSMR at two visits, multiple sclerosis, same patient20 (unverified)lesion-change mask onlyone imagedirectCC BY-NC-NDnon-commercialtruth to make
no truth on both images; Two raters place AFIDs on 10 visit pairs
Licence checked on 2026-10-10 at https://lit.fe.uni-lj.si/en/research/resources/3D-MR-MS/; derived test pairs shareable: unclear; commercial use: no. page 'License Creative Commons (CC): BY-NC-ND' (version not given); ND bars sharing adapted (resampled) pairs outside the teamTwo raters place AFIDs on 10 visit pairs; SynthSeg labels on both as a weaker check. The change mask is not a correspondence.
Longitudinal language processing in children 5-9 years (OpenNeuro ds003604)T1, DWI and fMRI of children at ages 5, 7 and 9, same subject ~2 years apart322 children; 184 with 2+ sessions, 69 with all 3 (T1 at each session unverified)nonedirectCC0opentruth to make
no truth on both images; AFIDs adapted for children by two raters on 15 pairs
Real growth deformation; one of few open child longitudinal sets.AFIDs adapted for children by two raters on 15 pairs; SynthSeg labels on both
Longitudinal Vestibular Schwannoma (TCIA 2025)MR at several visits, same patient534 scanstumour masksboth imagesdirectCC BY-NC-ND 4.0non-commercialin the corpuspriority 3
Same collection as VestibularSchwannomaMCRC2; only a growing tumour mask, no registration truth.
ND forbids sharing modified images
LPBA40, IBSR18, CUMC12, MGH10T1 MR, inter-subject40, 18, 12, 10 subjectsmanual labels, 56 to 128 regionsboth imagesaccountIBSR free for non-commercial use only; others not re-checkednon-commercialneeds an accountThe Klein 2009 evaluation sets. status change: usage counted from papers: LPBA40 (56 manual labels) is the most used manual-label brain set in these papers (Mindboggle-101 next, 9 papers); IBSR18, CUMC12, MGH10 appear almost only in Quicksilver-era work. status change: Round 4 count: listed in 12 of the 38 round-4 surveys and benchmark papers (Haskins 2020 survey; Fu 2020 review (PMB); Boveiri 2020 review; Zou 2022 review (Front Oncol); Hammoudeh 2023 survey; Chen X 2021 review (Prog Biomed Eng); Abbasi 2022 scoping review; Chen 2025 survey (MedIA); Azad 2023 ViT review; Han 2022 GAN registration review; Magic or Mirage 2024; rethink-reg 2024).
LUMIEREglioblastoma MR at many visits, same patientmanyautomatic segmentations (HD-GLIO, DeepBraTumIA)both imagesdirect: figshare collection 5904905, Imaging-v202211.zip (32.6 GB) readable by ranges, no accountCC0 (figshare items of the collection)openleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): open, but not among the six minor sets kept. It was proposed as: 91 glioblastoma patients with 638 study dates (T1, T1 post-contrast, T2, FLAIR as .nii.gz), with tumour masks from DeepBraTumIA and HD-GLIO-AUTO and brain masks, all automatic; expert RANO ratings are text, not images. Pairs of visits of one patient are a real longitudinal problem, but the tumour masks change by nature and are network outputs, so the corpus would make labels with a public network on both visits (automatic) and show the tumour masks as an unranked label set. CC0 on figshare, one 32.6 GB zip readable by ranges.
truth made by a program
LUMIR test setT1 MR, inter-subject590 test scanslabels and landmarks, kept by the organisersboth imageshiddenCC BY-NC-SA 3.0 (OpenBHB)non-commercialleft outIts training brains are in the corpus, with FastSurfer labels. status change: usage counted from papers; LUMIR 2025 also tests zero-shot on ADNI 1.5T/3T, NIMH (ds005752), UltraCortex 7T and T2/T2*/FLAIR images; LUMIR25 adds high-field and T1-T2 tracks. status change: Round 4 count: listed in 4 of the 38 round-4 surveys and benchmark papers (Chen 2025 survey (MedIA); Learn2Reg 2024 overview; LUMIR 2025; LUMirage 2025).
LUMIR25 (Learn2Reg 2025)T1 MR (OpenBHB), between subjects; 2025 edition>5,000 brains (train 51 GB, validation 430 MB)automatic labels; manual landmarks by a radiologist and a neurologist on a test subset (hidden)one imagehiddenrepository MIT; data under OpenBHB termsopenleft out
its images or its truth are not released
Listed in the Chen 2025 survey's repository. Validation labels coarse (10 classes); test truth on the server. Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Chen 2025 survey (MedIA)).Use the open OpenBHB images with FastSurfer labels on both (a weaker check, as for LUMIR in the corpus) and two raters placing AFIDs on 10 pairs.
M4Raw (low-field multi-contrast k-space)0.3 T brain MRI, T1w, T2w, FLAIR (+GRE) with repeated acquisitions, same subject same session158 train/val, 25 motion, 25 test (208 subjects); 2-6 repetitions per contrastnone; repetitions and contrasts are same-sessiondirectCC BY 4.0opentruth to make
no truth on both images; warp T2w by a known field and register T1w or FLAIR to it: exact truth after removing the small inter-scan motion (rigid pre-alignment); a later registration set must derive from M4Raw_T2_Syn, which holds 41 of these subjects (no subject in two datasets)
Paper reports measured inter-contrast motion; test set has about twice the motion of training. Repetitions give same-contrast pairs with tiny real motion; low field and low SNR make it a hard multimodal case.warp T2w by a known field and register T1w or FLAIR to it: exact truth after removing the small inter-scan motion (rigid pre-alignment); repeats give a noise-only control
Maclaren test-retest brain volume (OpenNeuro ds000239)T1 MR (ADNI protocol), 3 healthy adults, 20 sessions over 31 days, 2 scans per session with repositioning120 volumes (40 per subject): 60 intra-session pairs and many inter-session pairsnone (FreeSurfer volumes in the paper)directCC0 (OpenNeuro; also on figshare 10.6084/m9.figshare.929651)opentruth to make
no truth on both images; near-identity within sessions
Tiny and clean: ideal near-identity and transitivity check.near-identity within sessions; transitivity over 40 scans; AFIDs on a few pairs
MASiVar (OpenNeuro ds003416)Multisite, multiscanner diffusion MRI, repeat sessions97 subjectsNoneone imagedirectCC0opentruth to make
no truth on both images; Scan-rescan near identity and cycle consistency across sites
Scan-rescan; truth fails Licence checked on 2026-10-10 at https://openneuro.org/datasets/ds003416; derived test pairs shareable: yes; commercial use: yes. OpenNeuro snapshot 2.0.2 dataset_description License: CC0Scan-rescan near identity and cycle consistency across sites; two raters place 20 landmarks on b0/FA of 10 cross-site pairs (EPI distortion differs between scanners).
MCIC (MIND Clinical Imaging Consortium, schizophrenia)T1 brain MRI, between subjectsabout 330 subjects (unverified)none manualrequestMCIC Data Usage Agreement (SchizConnect/COINS): scientific investigation, teaching or planning clinical research; no disclosure beyond the approved application; team members bound by the DUA; no commercial clauseaskneeds a request
the data needs a request or a data use agreement
Fails truth. Distribution point (COINS, SchizConnect) not confirmed. Licence checked on 2026-10-10 at http://schizconnect.org/documentation/data-usage-agreements; derived test pairs shareable: unclear; commercial use: unclear. access via SchizConnect/COINS account and data use applicationlabels from a public segmenter on both, weaker check
MICA-MICs (microstructure-informed connectomics)3T T1w (two runs), quantitative T1 (MP2RAGE), multi-shell DWI, rs-fMRI; one session per subject50 healthy adultsnone (automatic parcellations)directCC0opentruth to make
no truth on both images; warp the T1 map or INV1 by a known field and register UNI or T1w to it: exact truth for MP2RAGE images
MP2RAGE gives UNI, INV1, INV2 and T1 map from one acquisition: aligned by construction. Second T1w run allows a repeat pair.warp the T1 map or INV1 by a known field and register UNI or T1w to it: exact truth for MP2RAGE images; T1w-qT1 across runs needs rigid pre-alignment
MICCAI 2012 Multi-Atlas LabelingT1 MR (OASIS), inter-subject30 scansNeuromorphometrics manual labelsboth imagesrequestCC BY-NCnon-commercialneeds a request
the data needs a request or a data use agreement
contact form
Midnight Scan Club (OpenNeuro ds000224)10 adults: 4 T1, 4 T2, 4 MR angiograms, 8 MR venograms over 2 sessions, plus 10 fMRI sessions; same subject10 subjects x (4 T1 + 4 T2) repeats, plus T1-MRA/MRV pairsnone for registration; hand-edited FreeSurfer surfaces from the averaged T1one imagedirectCC0opentruth to make
no truth on both images; vessel bifurcations found by a program on MRA/MRV then checked by a person
Angiograms and venograms of the same heads allow vessel-based checks.vessel bifurcations found by a program on MRA/MRV then checked by a person; T1-T1 repeats as near-identity check
Mind Brain Body (MPI-Leipzig LEMON)T1w, T2w and FLAIR brain MRI of young and old adults, between subjects226-228 subjectsnone manualdirectOpenNeuro (CC0 by default; not re-read)opentruth to make
no truth on both images; known smooth warp on the same-session T2w or FLAIR gives exact multimodal truth
Same-session multi-contrast scans are aligned by acquisition.known smooth warp on the same-session T2w or FLAIR gives exact multimodal truth
MIRIADLongitudinal T1 MR, 46 AD + 23 controls, same scanner, 2 weeks to 2 years, same-day back-to-back scans708 scans; back-to-back pairs and all interval pairsNone (no labels); back-to-back pairs give a near-identity checkone imageaccountregistration plus a MIRIAD data use agreement (screening note; the UCL page returns 403 to scripts, terms not re-read)askneeds an account
the download needs an account
Truth only by consistency (rescan = no change, transitivity). Licence checked on 2026-10-10 at https://www.ucl.ac.uk/drc/research/methods/miriad-scan-database; derived test pairs shareable: unclear; commercial use: unclear. MIRIAD host miriad.drc.ion.ucl.ac.uk returns an empty pageBack-to-back pairs as near identity, plus inverse and transitivity checks over all visits; two raters place AFIDs on 10 interval pairs.
MNI-HISUB253T submillimetre T1/T2 MR, 25 healthy adults; inter-subject25 subjects = 600 directed pairsManual hippocampal subfields (subiculum, CA1-3, CA4-DG) by expertsboth imagesaccountCC BY-NC; NITRC account and 1000FCP access requestnon-commercialneeds an account
the download needs an account
Hippocampus only.
Monash rsPET-MR, vis-fPET-fMRI, DaCRA (simultaneous PET/MR)simultaneous dynamic FDG PET, BOLD fMRI, T1w and field maps on a Siemens Biograph mMR, healthy adults27 (ds002898) + 10 (ds003382) + 15 (ds003397) subjectsnonedirectCC0opentruth to make
no truth on both images; warp T1w by a known field and register summed PET to it: exact multimodal truth after rigid head-motion correction of PET frames
Simultaneous acquisition: PET and MR share scanner geometry, so PET-T1w alignment is exact except head motion during the long PET. Dynamic PET frames also give a motion-correction problem.warp T1w by a known field and register summed PET to it: exact multimodal truth after rigid head-motion correction of PET frames
MR-ART motion-corrupted and clean brain MRIT1w with no, mild and strong head motion, same session148 subjects x 3 scans (unverified)noneone imagedirectCC0 (OpenNeuro, unverified)opentruth to make
no truth on both images; a known smooth warp on the clean scan gives truth
Pairs differ by rigid motion and artefact.a known smooth warp on the clean scan gives truth
MRBrainS13 and 18T1, T1-IR, FLAIR, same subject7 to 20manual tissue labelsboth imagesaccountMRBrainS13 data on DataverseNL (doi:10.34894/645ZIN) under 'Custom terms': 'See UMC Utrecht license' (licence text not attached; downloading implies acceptance); MRBrainS18 data page (archived) states no termsaskneeds an account
the download needs an account
Licence checked on 2026-10-10 at https://dataverse.nl/api/datasets/:persistentId/?persistentId=doi:10.34894/645ZIN; derived test pairs shareable: unclear; commercial use: unclear. challenge sites offline; during the challenges a confidentiality agreement was required; the referenced UMC Utrecht licence must be obtained
MRI to transcranial sonography co-registration database (Concordia, MNI)T1w MRI to 3D transcranial ultrasound, same healthy subject8 subjectsrigid transform averaged over 3 raters' manual registrations; inter- and intra-rater spread reportedboth imagesdirect: OSF project zdcjb, file by file, no accountcontradictory: OSF licence field 'CC-By Attribution 4.0 International' (2024); README and the file 'LICENSE-CC-BY-NC-SA 4.0' in the deposit: CC BY-NC-SA 4.0; ask the authors before commercial use Valentin's rule (2026-10-13): when a deposit contradicts itself, its LICENSE file, the stricter reading, wins: non-commercial.non-commercialleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): its licence is not shown to allow commercial use (contradictory: OSF licence field 'CC-By Attribution 4.0 International' (2024); README and the file 'LICENSE-CC-BY-NC-SA 4.0' in the deposit: CC BY-NC-SA 4.0; ask the authors before commercial use Valentin's rule (2026-10-13): when a deposit contradicts itself, its LICENSE file, the stricter reading, wins: non-commercial.). It was proposed as: OSF zdcjb (1.4 GB) holds 8 healthy adults, each a defaced T1w MRI and a left and a right transtemporal 3D ultrasound volume (15 cm depth) with an averaged rigid transform per side ('averaged transformation matrices', README), in NIfTI and MINC: 16 pairs from 8 subjects, under ten test pairs from distinct subjects. The OSF licence field says CC BY 4.0, but the deposit's README ('distributed under the CC BY-NC-SA 4.0 License') and its LICENSE file say CC BY-NC-SA 4.0. Brain US-MR is already in the corpus (RESECT, BITE) with expert landmarks.
First public MRI-TCS set; rigid only; also ships an MRI vessel atlas. Licence checked on 2026-10-10 at https://api.osf.io/v2/nodes/zdcjb/; derived test pairs shareable: yes; commercial use: yes. OSF project licence 'CC-By Attribution 4.0 International' (2024)two raters mark 10 vessel and ventricle points visible in both, to measure error beyond rigid
MRIxFields2026 (cross-field brain MRI, 0.1T to 7T)T1w, T2w and FLAIR brain MRI of the same travelling volunteers at 0.1, 1.5, 3, 5 and 7 T (plus 1,900+ unpaired scans)40 volunteers x 5 fields x 3 contrasts (3 volunteers in training; 17 validation and 20 test volunteers held out)none for registration; organisers rigidly then deformably (SyN) aligned each volunteer's fields to the 3T T1waccountunverified (challenge data terms)unknownneeds an account
the download needs an account
MICCAI 2026 synthesis challenge. Released pairs are pre-aligned; raw unaligned scans unverified. 0.1T to 7T is an extreme cross-field gap.a known smooth warp applied to the pre-aligned pairs gives exact cross-field truth; two people place 20 landmarks (AC, PC, vessel branch points) on 10 volunteers to check the provided alignment
MSSEG-2FLAIR at two visits, multiple sclerosis, same patient100new lesions on the second visit onlyone imageaccountOFSEP MRI Data Access Agreement (accepted on Shanoir-NG after challenge registration): no commercial use of the data or derivatives; no sharing except with people who signed the DUA themselves; delete after three yearsnon-commercialneeds an account
the download needs an account
Licence checked on 2026-10-10 at https://portal.fli-iam.irisa.fr/files/2021/03/OFSEP_MRI_Data_Access_Agreement_MICCAI_2021.pdf; derived test pairs shareable: unclear; commercial use: no. team sharing only if every member signs; derivative sharing only at group levelTwo raters place AFIDs on the FLAIR of 10 visit pairs; SynthSeg labels on both as a weaker check. New-lesion masks are not correspondences.
MTL subregion segmentation atlas with 7T and 3T multi-modality MRI (Dryad 2025)3T T1w/T2w and 7T MRI of the same subjects, medial temporal lobeatlas subjects with both field strengths (count unverified)MTL subregion segmentations (on which modality unverified)Dryad, free, no account for a person: the browser download sits behind an AWS 'confirm you are human' captcha and the Dryad API answers 401 without a token, so no scripted downloadCC0 1.0 (Dryad)openleft outpriority 2
25 subjects of the Penn ADRC ASHS atlas set: left and right MTL ROIs of the 7T T2w (0.42x0.42x1 mm) with ABC-protocol subregions (hippocampal subfields, ERC, BA35, BA36, PHC, sulci) drawn by hand, and the 7T T1w, 3T T2w and 3T T1w ROIs rigidly registered onto each 7T T2w ROI by the provider (Dryad README, Hippocampus 2025). The labels exist on the 7T T2w only and the 3T/7T images come aligned, so the scored task is between subjects: manual subregions on both ROIs, covering the crop. 25 subjects give 10 test pairs and 5 subjects (10 pairs among them) for development; synthesis 112 would hold the same ROIs. Left out on 2026-10-28: open (CC0) with hand-drawn subregions, but not among the six minor sets Valentin kept under the rule of 2026-10-27 (relevant sets only, no padding).
built to train multi-modality MTL segmentationif labels exist on both field strengths: label overlap; otherwise two people mark 10 hippocampal landmarks on 10 subjects at 3T and 7T
MU-Glioma-Postmulti-sequence MR, up to 6 post-operative visits, same patientup to 393network masks refined by neuroradiologists, tumour and cavity, every visitboth imagesdirect: IBM Aspera public link on the collection page (2,978 NIfTI files in patient and timepoint folders, which ascp fetches one by one as for UCSDBrainMets), no accountCC BY 4.0 on TCIA (images and segmentations, volumes, clinical data)openin the corpuspriority 3
In the corpus as segmentation; a registration set would need patients outside it and SynthSeg labels, a setting UCSDBrainMets already covers.
the tumour changes
Multicenter multi-shell diffusion of travelling adults (Tong 2020)T1 and multi-shell DWI of 3 adults on 10 identical Prisma scanners (one scanner 3 times)36 sessions (3 subjects x 12)nonedirectCC BY 4.0opentruth to make
no truth on both images; near-identity check
Same scanner model and protocol isolates the site effect.near-identity check; AFIDs on all 36 T1s
MUSHAC (CDMRI multi-shell diffusion harmonisation challenge)diffusion MR of the same people on different scanners~15 subjects (unverified)none for registrationrequestnot readable: CDMRI 2018 data page returns 403; no other host foundunknownneeds a request
the data needs a request or a data use agreement
Listed in Li 2023 (ACM CSUR) under registration; a harmonisation set, not a registration benchmark. Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Li 2023 dataset collection (ACM CSUR)). Licence checked on 2026-10-10 at https://projects.iq.harvard.edu/cdmri2018/data; derived test pairs shareable: unclear; commercial use: unclear. harmonisation challenge data (Cardiff/CUBRIC scans), access terms unknownsame-subject cross-scanner pairs: register T1 of each session; labels from SynthSeg (weak)
MyConnectomeOne subject, ~100 sessions over 18 months (structural subset)Repeat sessionsNonedirectCC0 (OpenNeuro)opentruth to make
no truth on both images; Repeat sessions of one subject: near-identity, inverse and cycle consistency only
Fails truth.Repeat sessions of one subject: near-identity, inverse and cycle consistency only; one brain cannot make a test set.
Natural Scenes Dataset (NSD)7T fMRI plus anatomy (T1, T2, DWI, venogram, angiogram) of 8 adults over 30-40 sessions8 subjects; repeated T1/T2 and 7T sessions (counts unverified)nonerequestNSD Data Access Agreement form (terms given in the form, not on the site)askneeds a request
the data needs a request or a data use agreement
Angiograms and venograms of the same heads. Licence checked on 2026-10-10 at https://naturalscenesdataset.org/; derived test pairs shareable: unclear; commercial use: unclear. after the form, a link to the data manual and anonymous S3 accessvessel bifurcations on angiogram/venogram by program then checked by a person; T1 repeats as near-identity
NeoBrainS12Preterm neonate T1/T2 MR at 30 and 40 weeks; inter-subject (same infants at 30 and 40 weeks: unverified)Few training scans per set (3 sets)Manual labels of 8 tissue classes (UMC Utrecht)both imagesaccountnot readable: neobrains12.isi.uu.nl offline, archived snapshot redirects to the ISI home page; no DataverseNL record foundunknownneeds an account
the download needs an account
Site did not answer on 2026-10-10. If the same infants exist at 30 and 40 weeks, it is a rare longitudinal neonatal set. Licence checked on 2026-10-10 at http://web.archive.org/web/20230205094105/https://neobrains12.isi.uu.nl/; derived test pairs shareable: unclear; commercial use: unclear. ask UMC Utrecht (ISI) for terms
NeuriPhy (physics-informed brain shift dataset)Pre-op brain MRI with simulated brain shift; ReMIND + UPENN-GBM207 cases (45 ReMIND, 162 UPENN-GBM)Synthetic biomechanical displacement fields, segmentations, keypointsboth imagesrequestCC BY 4.0 (record restricted)openneeds a request
the data needs a request or a data use agreement
Work in progress; files not public yet
NeuroEXPLORER vs HRRT paired brain PET with T1w (OpenNeuro ds006917)same subjects scanned with the same tracer on two PET scanners (0.5 mm and 1.2 mm voxels), 7 tracers, plus T1w MRI7 subjects, NX-HRRT pairs per tracernoneone imagedirectOpenNeuro (CC0 typical, unverified)opentruth to make
no truth on both images; T1w-based rigid alignment of each PET checked by a person on gyral landmarks
PET-PET across scanners and PET-MR; very high resolution PETT1w-based rigid alignment of each PET checked by a person on gyral landmarks; or a known warp on one PET
Neuromorphometrics databaseT1 MR, inter-subjecthundredsmanual labelsboth imagesrequestpaid subscription: academic subscription or commercial data set licenceaskneeds a request
the data needs a request or a data use agreement
Licence checked on 2026-10-10 at https://www.neuromorphometrics.com/; derived test pairs shareable: no; commercial use: unclear. commercial licence is sold; terms of each licence not public
NextBrain (Casamitjana 2025, BUNGEE-TOOLS)Ex vivo MRI vs ~10,000 histology sections and blockface photos, 5 human hemispheres; same specimen and inter-subject5 hemispheres (MRI-histology); 20 directed inter-hemisphere pairsManual delineations of 333 regions on histology sections; aligned version from the authors' own registrationone imagedirectCC0 (figshare API); extension ds005422 CC0opentruth to make
no truth on both images; Two people place 10 landmarks (vessels, region boundaries) on 10 histology section/MRI slice pairs per hemisphere, using the blockface photos to find the slice
Top pick for MR-histology: open, huge, expert labels. Labels on histology only.Two people place 10 landmarks (vessels, region boundaries) on 10 histology section/MRI slice pairs per hemisphere, using the blockface photos to find the slice; the authors' aligned version is a reference only.
Nigrosome atlas from 3D histologypost-mortem MRI and 3D histology of the substantia nigra, warped to a templatebrain count unverifiednigrosome probability maps and registration scriptsone imagedirectCC BY 4.0opentruth to make
no truth on both images; label nigrosomes on paired 7T scans as a target
Atlas, no paired truth. Licence checked on 2026-10-10 at https://api.osf.io/v2/nodes/gsphy/; derived test pairs shareable: yes; commercial use: yes. OSF project licence 'CC-By Attribution 4.0 International' (2025)label nigrosomes on paired 7T scans as a target
NIMH Healthy Research Volunteer Dataset (OpenNeuro ds005752)T1, T2, DWI, fMRI and MEG of healthy adults; few re-evaluation visits251 subjects in snapshot 2.1.0 (sessions 01/02); 15 re-evaluations in v2.0.0; MRI repeat count unverifiednonedirectCC0opentruth to make
no truth on both images; between subjects: AFIDs or SynthSeg labels on both
Mostly a between-subject pool; same-subject repeats are few. status change: Round 4 count: listed in 2 of the 38 round-4 surveys and benchmark papers (LUMIR 2025; LUMirage 2025).between subjects: AFIDs or SynthSeg labels on both
NIREP NA0T1 MR, 16 adults (8 M, 8 F), rigidly aligned; inter-subject16 = 240 directed pairsManual labels of 32 grey-matter regionsboth imagesaccountNITRC page: free for non-commercial use onlynon-commercialneeds an account
the download needs an account
Status change risk: nirep.org now hosts a gambling site; whether NITRC still serves NA0 files is unverified.
NKI-Rockland Sample (Enhanced NKI-RS)multimodal brain MR, community sample aged 6-85, some studies with follow-up visits>1,000 participants; number with repeat imaging visits unverified (count from the AWS links file)nonedirectINDI terms: after NITRC registration, 'unrestricted usage of the datasets for non-commercial purposes'; imaging release needs no DUA (full phenotype needs a notarised DUA)non-commercialtruth to make
no truth on both images; two raters place the 32 AFIDs fiducials (published protocol) on both scans of 20 pairs
Anonymous S3 (fcp-indi/data/Projects/RocklandSample/RawDataBIDSLatest). Licence checked on 2026-10-10 at http://fcon_1000.projects.nitrc.org/; derived test pairs shareable: unclear; commercial use: no. imaging on the public fcp-indi S3 bucket; the NKI-RS access page itself names no licencetwo raters place the 32 AFIDs fiducials (published protocol) on both scans of 20 pairs; SynthSeg labels on both
NRM2018 PET Grand Challenge (simulated dynamic PET with MR)simulated 90-min dynamic PET of a fictional tracer, baseline and displaced scans, with the subject's MR5 simulated subjects x 2 scansknown ground truth of binding change (simulation); PET built on the MR anatomyboth imagesdirectCC0opentruth to make
5 simulated subjects whose PET is built on the MR anatomy, so only a known warp would make a truth.
PET is simulated on the MR anatomy so PET-MR alignment is exact by construction; low noise realism unverified.warp MR by a known field and register PET frames to it: exact truth; vary frame (early vs late) to change contrast
NYUMets_Brainlongitudinal brain MRI (T1, T1 post-contrast, T2, FLAIR) of patients treated with Gamma Knife radiosurgery, same patient1,429 patients, 8,003 studies, median 6 time points2,367 expert tumour segmentations from radiosurgery planningone imagecontrollednon-commercial use only; request form and AWS account (ARN) requirednon-commercialneeds a request
the data needs a request or a data use agreement
Gamma Knife planning MRIs are usually frame-based (Leksell); whether the fiducial box survives the 1 mm resampling and preprocessing is unverified.on planning-day MRIs, the Leksell indicator box N-marks give exact frame coordinates per slice; between sessions the frame changes, so two people place AFIDs on 20 pairs; tumour masks give a weak check
OASIS-2T1 MR, adults 60-96 with and without dementia, same subject over 1+ years150 subjects, 373 sessions (2-5 visits, at least 1 year apart); 3-4 T1 scans per sessionnone (clinical ratings only)directOASIS Data Use Terms: non-commercial academic research onlynon-commercialtruth to make
no truth on both images; two raters place the 32 AFIDs fiducials (published protocol) on both scans of 20 pairs
Same-session repeat T1s give exact rescan pairs plus 1-5 year pairs; page has direct archive links (login need unverified). OASIS-3 is listed.two raters place the 32 AFIDs fiducials (published protocol) on both scans of 20 pairs; same-session repeats as near-identity check
OASIS-3Longitudinal T1/T2/FLAIR MR and PET, ageing and dementia; same subject over years~1,400 participants, ~2,800 MR sessions (unverified)FreeSurfer outputs (automatic)both imagesaccountOASIS Data Use Agreement: non-commercial academic research only (commercial use incl. consulting prohibited without written authorisation); no sharing with any third party, collaborators included, without written permissionnon-commercialneeds an account
the download needs an account
Fails expert truth. status change: Round 4 count: listed in 2 of the 38 round-4 surveys and benchmark papers (Abbasi 2022 scoping review; Han 2022 GAN registration review). Licence checked on 2026-10-10 at https://sites.wustl.edu/oasisbrains/home/access/; derived test pairs shareable: no; commercial use: no. every collaborator must apply separately; data on secure access-controlled systems only
ON-HarmonyT1/T2/dMRI/SWI/rsfMRI, 20 subjects on 6 scanners (3 vendors), ~half with 5 within-scanner repeats165 sessions; many same-subject pairsNonedirectCC0 (OpenNeuro)opentruth to make
no truth on both images; Within-scanner repeats as near identity, cross-scanner pairs by cycle consistency
Best open travelling-heads set; truth by consistency only.Within-scanner repeats as near identity, cross-scanner pairs by cycle consistency; two raters place AFIDs on 10 cross-vendor pairs; same-session T1/T2 under a known warp give near-exact multi-contrast truth.
OpenBHB (Open Big Healthy Brains)T1 brain MRI of healthy controls from 10 public studies, between subjectsabout 3,984 scans used by LUMIRnone manual (quasi-raw and preprocessed images)directCC BY-NC-SA 3.0 (as stated for LUMIR)non-commercialtruth to make
no truth on both images; FastSurfer labels on both (as LUMIR training already has), weaker check
Training images of LUMIR; truth exists only on the hidden LUMIR test set.FastSurfer labels on both (as LUMIR training already has), weaker check; AFIDs by two raters on 20 subjects
OpenNeuro brain PET with repeated PET sessions and T1w (ds007768, ds004733, ds004731, ds004730, ds004654, ds005698, ds007561, ds006756)dynamic PET scans repeated within or across sessions (FDG crossover, leucine sleep/wake, test-retest PDE4B, UCB-J, MK6240) plus T1w MRI, same subjectabout 200 subjects in total, 2-3 PET scans eachnone; some sets ship MRI already co-registered to PET by the authorsone imagedirectOpenNeuro (CC0 typical, unverified)opentruth to make
no truth on both images; cerebellum/cortex labels from FreeSurfer on T1w transferred to each PET
PET-PET between scans (head repositioning) and frame-to-frame motion; authors' co-registrations are not independentcerebellum/cortex labels from FreeSurfer on T1w transferred to each PET; a known rigid motion applied to one scan for exact truth
Other OpenNeuro simultaneous PET/MR brain sets (ds004513, ds008786, ds007907, ds007135)Biograph mMR PET/MR: FDG-PET + BOLD + DWI (TUM, ds004513, ds008786); [11C]PBR28 PET + T1w (MGH, ds007907); Martinostat PET + T1w/FLAIR/BOLD (ds007135)20 + 15 + 147 + 25 subjectsnonedirectCC0opentruth to make
no truth on both images; warp the T1w by a known field and register PET to it: exact truth up to PET head motion
ds007907 T1w is already coregistered to PET by the authors. Different tracers give different PET contrast: useful to vary the multimodal difficulty.warp the T1w by a known field and register PET to it: exact truth up to PET head motion
Paired 3T/7T MRI with manual hippocampal subfields (Chu 2025)3T and 7T T1w/T2w/rsfMRI of the same 20 volunteers; cross-field same subject20 pairs (3T-7T), plus T1-T2Manual hippocampal subfields (7 regions) on 7T T2w onlyone imagedirectCC BY 4.0 (figshare API)opentruth to make
no truth on both images; Two raters draw the hippocampal subfields on 3T T2w of 10 subjects with the 7T protocol: contours on both
Labels on one side only; can score 7T-to-3T label transfer only with labels mapped by the method.Two raters draw the hippocampal subfields on 3T T2w of 10 subjects with the 7T protocol: contours on both; AFIDs on 3T and 7T T1w for point truth.
Paired 64 mT and 3T brain MRI (Zenodo)64 mT and 3T T1, T2, FLAIR and DWI, same healthy subject11 paired subjects (86 low-field scans)noneone imagedirectMIT Licenseopentruth to make
no truth on both images; SynthSeg labels on both (weaker)
Low-field to high-field.SynthSeg labels on both (weaker); two people place 20 landmarks on 11 pairs
Paired CT and MRI dataset (Dhaka, Data in Brief 2025)CT with T1 and T2 MRI, same patient (anatomy unverified, likely head)18 patientsco-registration confirmed by radiologists, no labelsone imagedirectCC BY-NC (article; data unverified)non-commercialtruth to make
no truth on both images; two people mark 10 landmarks on 10 pairs
Hosted on Kaggle.two people mark 10 landmarks on 10 pairs
Pediatric brain MRI with and without motion correction (PublicnEUro)clinical brain MRI of children scanned without anaesthesia, with and without markerless prospective motion correction, same session47 children, repeated sequenceshead motion traces from the markerless tracker; radiologist quality ratingsaccountunverified (PublicnEUro terms)unknownneeds an account
the download needs an account
Sci Data 2026. Motion traces give head pose during each scan, not a full scan-to-scan transform.use tracker poses at scan midpoints as rigid truth between repeats, checked with 10 landmarks per child on 10 children
Penn epilepsy paired 3T and 7T MRI with iEEG3T and 7T T1w, T2w, FLAIR and rs-fMRI, same epilepsy patient30 patients (15 with iEEG)noneone imagedirectCC BY-NC-ND 4.0 (article; data unverified)non-commercialtruth to make
no truth on both images; electrode contacts on post-implant imaging as fiducials
Repository not checked.electrode contacts on post-implant imaging as fiducials
Penn recurrent brain tumour database (Ou 2014)T1 MR of recurrent brain tumour with resection cavity registered to a healthy template, between subjects8 patients to 1 templatelandmarks and regions by two independent expertsboth imageshiddennot releasedaskleft out
its images or its truth are not released
Not public.request from Penn CBICA; BraTS-Reg is the public analogue
PPMI (Parkinson's Progression Markers Initiative)T1 (DTI subset) and DaTSCAN SPECT, Parkinson's and controls, same subject yearlythousands of participants with repeat MRI (count unverified)noneone imagecontrolledPPMI Data Use Agreement (no redistribution; restricts AI tools without data containment)askneeds a request
the data needs a request or a data use agreement
DUA forbids sharing participant-level data: a benchmark could link to it, not host it. status change: usage counted from papers: training pool, and T1-dMRI pairs in 2026 work. status change: Round 4 count: listed in 6 of the 38 round-4 surveys and benchmark papers (Fu 2020 review (PMB); Boveiri 2020 review; Bharati 2022 review; Zou 2022 review (Front Oncol); Hammoudeh 2023 survey; Chen X 2021 review (Prog Biomed Eng)).two raters place the 32 AFIDs fiducials (published protocol) on both scans of 20 pairs (includes brainstem points relevant to Parkinson's)
Pre- and post-surgery brain tumour MRI (Ghent)T1, DWI and BOLD before and 6 months after tumour resection, same patient25 patients plus 11 controls (post-op in ds002080)pre-operative tumour masks onlyone imagedirectCC0 (OpenNeuro, unverified)opentruth to make
no truth on both images; two people place 15 landmarks on 10 pre/post pairs
Fails truth on both images.two people place 15 landmarks on 10 pre/post pairs
Pre- and post-surgical MRI in chronic trigeminal neuralgia (OpenNeuro ds005713)3D T1w, 3D T2w, DTI and rs-fMRI before and after microvascular decompression, same patient64 patients with a follow-up MRI (119 patients, 53 controls in all)nonedirectCC0opentruth to make
no truth on both images; two raters place the 32 AFIDs on 10 pre/post pairs
found through the OpenNeuro GraphQL APItwo raters place the 32 AFIDs on 10 pre/post pairs; SynthSeg labels on both as a weaker check
Pre- and postoperative multi-sequence MRI-CT glioma dataset (Zenodo 19198942)glioma MRI and CT before and after surgerycount unverifiedunverified (radiomics features file)directCC BY 4.0opentruth to make
pre- and postoperative MRI and CT in a nested zip with no description; no mask or landmark file is listed, only radiomics features (h5)
Chinese file names, no description.two people place 15 landmarks on 10 pairs
PREVENT-AD open dataT1, T2, DWI, fMRI of cognitively normal adults at risk of Alzheimer's, same subject yearly308 participants, up to 5 years of yearly imagingnoneaccountOpen PREVENT-AD Terms of Use: 'neurosciences research' only (per consent); do not sell or claim IP in the dataset; do not redistribute, share only through the portalaskneeds an account
the download needs an account
Account requested from the maintainers; registered set (348) adds clinical data. Licence checked on 2026-10-10 at https://openpreventad.loris.ca/images/Open_PREVENT-AD_Terms_of_Use.png; derived test pairs shareable: no; commercial use: unclear. terms image linked as the licence in the CONP DATS.json; registration on the LORIS/CONP portaltwo raters place the 32 AFIDs fiducials (published protocol) on both scans of 20 pairs (baseline vs last visit); SynthSeg labels on both
QIN GBM Treatment ResponseNewly diagnosed glioblastoma: "double baseline" multi-parametric MRI (test-retest)54 patients, 2 baselines eachNoneone imagecontrolledNIH Controlled Data Access Policyaskneeds a request
the data needs a request or a data use agreement
Fails truth. Images moved to dbGaP controlled access (phs004225, head images) in July 2025.Double baselines days apart: near-identity and inverse consistency; two raters place AFIDs on 10 pairs; same-session sequences under a known warp give near-exact multi-contrast truth.
ReMINDpre-op MR, intra-op MR, intra-op ultrasound, same patient114 patientspre-op tumour, intra-op residual, cavity on ultrasound: different structures, no landmarksone imagedirectCC BY 4.0opentruth to make
no truth on both images; Two people place 10 landmarks (vessels, sulci, ventricle edges) on pre-op MR and pre-resection iUS of 10 patients (RESECT protocol)
status change: Round 4 count: listed in 2 of the 38 round-4 surveys and benchmark papers (Learn2Reg 2024 overview; ReMIND2Reg 2025).Two people place 10 landmarks (vessels, sulci, ventricle edges) on pre-op MR and pre-resection iUS of 10 patients (RESECT protocol); for pre-op vs intra-op MR, AFIDs away from the resection.
ReMIND reference segmentations (ioUS-to-MR synthesis benchmark)Pre-op MRI (T2/FLAIR) vs intra-op US, pre- and post-resection, same patient297 label maps, 77 subjects (150 MR-space, 147 US-space)Tumour and resection cavity, semi-automatic + neurosurgeon reviewboth imagesdirect: Zenodo (labels, RAR); the images are TCIA's ReMIND, open through NBIAMIT in the Zenodo licence field; the record's description and the code's DATA.md say CC BY 4.0, as ReMINDopensame as
Labels only (a 1 MB RAR, 297 maps for 77 ReMIND subjects, made in ITK-SNAP with nnInteractive and reviewed by a neurosurgeon). Both label sets sit on the paper's pre-processed grid: each iUS rigidly registered to the MR in ImFusion (LC2), resampled onto the MR grid and cropped to the iUS field of view; neither those volumes nor the transforms are released, and the US labels were drawn MR-assisted (build_dataset_us.py), so no independent US truth lands on ReMIND's own iUS. ReMIND's registration row (375) keeps its status.
status change for ReMIND: new labels on both modalities (MR labels drawn after co-registration to US)
ReMIND2Regpre-op MR to post-resection ultrasound, same patient10 testlandmarks, kept by the organisersboth imageshiddenCC BY 4.0 (ReMIND images)openleft outTraining data come without annotations. status change: usage counted from papers. status change: Round 4 count: listed in 3 of the 38 round-4 surveys and benchmark papers (Chen 2025 survey (MedIA); Learn2Reg 2024 overview; ReMIND2Reg 2025).
Rennes 18-subject brain MR database (Hellier 2003)T1 MR of healthy subjects, between subjects18 subjectsglobal tissue overlap and local cortical-sulci measures (details unverified)hiddennot releasedaskleft out
its images or its truth are not released
Not public; later work moved to LPBA40 and Mindboggle.request; labels from a public segmenter, a weaker check
RFUds (brain metastases radiotherapy follow-up)Planning CT + MRI and follow-up MRIs, same patient44 patients, multiple follow-upsRT structures delineated at planning only; CT-MRI matrices providedone imagedirectCC BY 4.0opentruth to make
no truth on both images; CLAHE alters intensities but keeps geometry
Images contrast-processed (CLAHE)CLAHE alters intensities but keeps geometry; two raters place AFIDs on 10 planning/follow-up MRI pairs. The CT-MRI matrices are clinical references, not truth.
RHUH-GBMPre-op, post-op and follow-up GBM MRI (unverified)40 patientsTumour segmentations (unverified)both imagesaccountCC BY 4.0openneeds an account
the download needs an account
Public part is skull-stripped and each timepoint registered to SRI24, so pairs start roughly aligned; resection changes anatomy. Raw head DICOM moved to dbGaP in July 2025. Licence checked on 2026-10-10 at https://www.cancerimagingarchive.net/collection/rhuh-gbm/; derived test pairs shareable: yes; commercial use: yes. TCIA table now Public: NIfTI images + segmentations CC BY 4.0 (row said Limited)
RIDER Neuro MRIRepeat brain MRI of glioma patients19 patientsNone (unverified)one imageaccountNIH Controlled Data Access Policyaskneeds an account
the download needs an account
Scan-rescan Licence checked on 2026-10-10 at https://www.cancerimagingarchive.net/collection/rider-neuro-mri/; derived test pairs shareable: unclear; commercial use: unclear. TCIA Limited accessScan-rescan near identity and inverse consistency; AFIDs by two raters on 10 pairs for point truth.
RIRECT, MR, PET of the same patient18fiducial-based transforms, kept hiddenhiddenunverifiedunknownleft out
its images or its truth are not released
evaluation service no longer maintained status change: Round 4 count: listed in 4 of the 38 round-4 surveys and benchmark papers (Boveiri 2020 review; Bharati 2022 review; Chen X 2021 review (Prog Biomed Eng); Han 2022 GAN registration review).The fiducial-based transforms exist but are hidden: ask the owners to release them; otherwise two people place 15 landmarks on CT, MR and PET of the test patients.
sEEG Forced Two-Choice Task (OpenNeuro ds004473)T1w and T2w MRI and CT with sEEG depth electrodes, same patient; epilepsy8 subjects (MR-CT)contact positions in ACPC and scanner RAS, from the authors' CT-MR co-registration and markingone imagedirectCC0opentruth to make
no truth on both images; contact centroids on CT by threshold
Whether the T1 is pre- or post-implant is unverified. Contact positions are not independent of the authors' registration.contact centroids on CT by threshold; if T1 is post-implant, contact artefacts give matching points; otherwise two people place ~15 landmarks (AC, PC, vessels, skull) on T1 and CT; limit: metal blooming on CT
SIMONOne man, 73 sessions on 36 scanners over 15+ years, T1 every session plus other contrasts73 sessionsNonedirectCC BY-SA 4.0opentruth to make
no truth on both images; One subject on many scanners: cross-scanner near identity and cycle consistency
Fails truth.One subject on many scanners: cross-scanner near identity and cycle consistency; AFIDs by two raters on 10 sessions. One brain is a check, not a test set.
SLIM (Southwest University Longitudinal Imaging Multimodal)T1, DWI, rs-fMRI of young adults, same subject over ~3.5 years580/240/228 subjects at 3 time points; 121 with all 3, 211 with 2nonedirectCC BY-NC (Attribution - Non-Commercial)non-commercialtruth to make
no truth on both images; two raters place the 32 AFIDs fiducials (published protocol) on both scans of 20 pairs
Mean 818 days first-to-third scan; little anatomical change, so a near-identity longitudinal test. Anonymous S3.two raters place the 32 AFIDs fiducials (published protocol) on both scans of 20 pairs; transitivity over 3 time points
SPINS Traveling Human Phantoms (OpenNeuro ds003011)T1, DWI, rs-fMRI of 4 healthy men scanned yearly on 6 MRIs at 3 sites over 4 years30 sessionsnonedirectCC0opentruth to make
no truth on both images; cross-scanner near-identity
Scanner upgrades mid-study (GE and Trio to Prisma).cross-scanner near-identity; AFIDs on all 30 sessions (small enough)
SR-Reg (MambaMorph brain MR-CT)brain MRI to CT, same patient, derived from SynthRAD2023180 subjectssegmentation labels on MR and CT (method unverified, likely automatic)both imagesdirectdata derived from SynthRAD2023, which is CC BY-NC 4.0 (Zenodo 7260705); repos MIT for codenon-commercialleft out
its images are SynthRAD 2023 brain MR and CT, in the corpus; its labels could join SynthRAD_BrainMRCT
Derived from SynthRAD_BrainMRCT already in the corpus; adds labels. Licence checked on 2026-10-10 at https://zenodo.org/api/records/7260705; derived test pairs shareable: yes; commercial use: no. SR-Reg itself (Google/Baidu drive via mileswyn/MambaMorph) states no data licence; inherits NC from SynthRAD2023expert labels on 10 test pairs
SRPBS Traveling SubjectsT1 MR and rsfMRI, 9 men scanned at 12 sites; same subject across scanners411 sessionsNoneaccountDecNef Brain Data Repository terms: no commercial use; no copying or redistribution of the dataset; acknowledgement requirednon-commercialneeds an account
the download needs an account
Fails truth; cross-scanner near-identity check only. Licence checked on 2026-10-10 at https://bicr-resource.atr.jp/srpbsts/; derived test pairs shareable: no; commercial use: no. registration form issues a loginCross-site near identity and cycle consistency; AFIDs by two raters on 10 cross-site pairs.
St Olav/SINTEF intraoperative ultrasound brain shift cases (Iversen 2018)preoperative MR to navigated intraoperative 3D ultrasound, tumour resection, same patient13 casesmanually placed anatomic landmarksboth imageshiddennot releasedaskleft out
its images or its truth are not released
Not public; RESECT and ReMIND are the public analogues.request from SINTEF/NTNU
STBI serial head CT (Shiraz)serial non-contrast head CT after traumatic brain injury, same patient3,602 patients, 12,155 CTcategorical findings per scan (midline shift, haematoma), no masks statedone imagedirectOpen Data Commons Attribution License (ODC-By 1.0)opentruth to make
no truth on both images; public CT brain segmenter on both scans (weaker) plus 15 landmarks by two people on 20 pairs
Iranian data on the Chinese Science Data Bank; huge longitudinal pool. Licence checked on 2026-10-10 at https://api.datacite.org/dois/10.57760/sciencedb.38999; derived test pairs shareable: yes; commercial use: yes. ScienceDB record rights URI opendatacommons.org/licenses/by/1-0/public CT brain segmenter on both scans (weaker) plus 15 landmarks by two people on 20 pairs
studyforrest (OpenNeuro ds000113)3T and 7T MR of 20 adults: T1, T2, SWI, angiography, DTI, 7T fMRI; same subject multimodal20 subjects (+ phantom), several sessions and field strengthsnone (pre-computed within-subject alignments shipped, not truth)directODC Public Domain Dedication and Licence (PDDL)opentruth to make
no truth on both images; same-session contrasts give multimodal truth under a known warp
3T vs 7T of the same head and several contrasts. Licence checked on 2026-10-10 at https://www.studyforrest.org/access.html; derived test pairs shareable: yes; commercial use: yes. studyforrest.org Terms of Use: 'All data are released to the public under the PDDL', redistribution encouraged with a reference; OpenNeuro snapshot License field emptysame-session contrasts give multimodal truth under a known warp; vessel bifurcations on angiography for 3T-7T pairs
Tohoku Medical Megabank Brain MRI Study3D T1w, FLAIR, DTI, MRA, ASL of the same adults about 5 years apart12,164 at baseline; follow-up started 2019 (count unverified)nonecontrolledTMM Biobank data accessaskneeds a request
the data needs a request or a data use agreement
URL is the design paper; shared through the TMM BiobankSynthSeg labels on both visits as a weak check; AFIDs by two raters on 10 visit pairs
Trigeminal neuralgia MRI-MRA benchmarkStructural MRI vs TOF-MRA, same patient149 patientsClinician-annotated bilateral trigeminal ROIs, partial vessel annotationshiddenUnverifiedunknownleft out
its images or its truth are not released
Release not statedTwo people place 10 landmarks (basilar tip, SCA and AICA origins, nerve root entry) on structural MRI and TOF-MRA of 10 patients; COSTA vessel labels on the MRA as a weaker check.
Trigeminal neuralgia pre- and post-surgical MRI (OpenNeuro)structural (T1w, T2w) and functional MRI before and 6 months after surgery, same patient58 patients with follow-up (112 patients, 48 controls)noneone imagedirectunverified (OpenNeuro)unknowntruth to make
no truth on both images; subcortical labels from a public segmenter on both visits (weaker)
Sci Data 2025. Different from the JHU MRI-MRA benchmark already listed.subcortical labels from a public segmenter on both visits (weaker); two people place 15 landmarks on 10 pairs
UCSD-PTGBMPost-treatment high-grade glioma MRI (incl. diffusion RSI, perfusion), some patients at more than one timepoint178 patients, 243 timepoints (at most ~65 longitudinal pairs)BraTS-style tumour labels per timepoint, manually correctedboth imagesdirectCC BY 4.0opentruth to make
every visit is resampled to the MNI atlas (pairs already registered); the truth is the changing tumour.
All modalities resampled to MNI atlas, so pairs are pre-aligned; part of BraTS 2024 post-treatment test data.
UCSD-VS-LongitudinalMRI T1-pre, T1-post, T2 (+ IAC, FLAIR, DWI, SWI, ASL) of vestibular schwannoma, same patient at 2-4 timepoints191 patients, 570 timepoints (about 380 pairs)Tumour mask on T1-post at every timepoint (nnU-Net then manual correction, neuroradiologist-approved); growth class per subjectboth imagespublic Aspera (Faspex) link on TCIA, no account, file by file (4,391 files)CC BY 4.0 (collection page)openleft outpriority 3
Built and reviewed on 2026-10-14, then left out by Valentin: the truth is SynthSeg run by the corpus on each image, eight of ten test pairs have 4.5 to 5 mm slices, and between two MRIs of one adult head rigidly aligned there is almost nothing to deform (no registration 0.70 Dice, ConvexAdam 0.67 and lower on 8 of 10 pairs). The segmentation set of the same collection, VestibularSchwannomaUCSD_Seg, is in the corpus.
Published 2026-04. Native space, not registered, defaced. Small tumour labels only: weak global check. Different cohort from VS-MC-RC / MC-RC2 (King's); MC-RC2 (TCIA 2025) is probably the "Longitudinal VS" already covered.
UCSF-ALPTDGpost-treatment glioma MR at two visits, same patient298expert-corrected tumour masksboth imagesrequestnon-commercial data use agreementnon-commercialneeds a request
the data needs a request or a data use agreement
visits already registered to visit 2 and to SRI24
UTSW-Glioma (TCIA 2026)pre-surgical glioma MRI (pre/post-contrast T1, T2, FLAIR), same session, multi-vendor; between patients625 patientsmulti-label tumour segmentationsboth imagesdirectunverified (TCIA collection licence)unknowntruth to make
one time point with aligned contrasts, so only a known warp; between patients the truth is the tumour alone; licence not stated.
Single time point; multi-contrast is the usable pairing.known smooth warps on the aligned contrasts for exact truth; public segmenter labels on both images for inter-subject (weaker)
Vanderbilt model-updated brain shift cases (Miga 2016)preoperative MR to intraoperative cortical surface deformation data, tumour resection, same patient16 patientssparse intraoperative cortical surface deformation (acquisition method unverified)one imagehiddennot releasedaskleft out
its images or its truth are not released
Sparse surface truth only, not volumetric; not public.request from Vanderbilt; iUS-to-MR landmarks by two people if intraoperative ultrasound exists
Vestibular-Schwannoma-MC-RCT1c and T2 at several visits, same patientup to 268manual tumour mask at every visitboth imagesdirect: TCIA NBIA API, no account; masks as a zip on TCIA's siteCC BY 4.0 (collection page and every NBIA series)openin the corpuspriority 3
In the corpus as segmentation; one growing tumour, no registration truth.
small target that grows
White matter multi-scale data: diffusion MRI and synchrotron X-ray (Zenodo 10458911)diffusion MRI vs synchrotron X-ray scans of the same white matter samples (vervet monkey, healthy human, MS)count unverifiednonedirectCC BY 4.0opentruth to make
no truth on both images; vessel landmarks in both scales marked by two people
vessel landmarks in both scales marked by two people
Whole-brain 3D MRF + MRI scan-rescan (3 scanners)5 subjects x 3 scanners x 2 scans30 scansNoneone imagedirectCWRU EULA: non-commercial academic research by non-profit institutions onlynon-commercialtruth to make
no truth on both images; MRF T1, T2 and PD maps come from one acquisition and are aligned by construction: a known warp gives exact multi-contrast truth
Scan-rescan; truth fails Licence checked on 2026-10-10 at https://zenodo.org/records/8234101/files/LICENSE.pdf; derived test pairs shareable: no; commercial use: no. EULA forbids copying, derivative works, distribution and service-bureau use without CWRU written consentMRF T1, T2 and PD maps come from one acquisition and are aligned by construction: a known warp gives exact multi-contrast truth; scan-rescan and cross-scanner pairs give near-identity checks.
Yale brain metastases longitudinalMR at several visits, same patientmanynonedirectCC BY 4.0opentruth to make
no truth on both images; Two raters place AFIDs away from lesions on 10 visit pairs
no segmentations (link to verify)Two raters place AFIDs away from lesions on 10 visit pairs; SynthSeg labels on both as a weaker check.

Head and neck · 47

DatasetProblemPairsChecked againstAccessLicenceStatusNotesTo make the truth
3D cephalometric CBCT landmarks (Aalto/Oulu, PLOS ONE 2024)CBCT skull, multi-centre (Finland, Thailand); inter-subject309 CBCT46 cephalometric landmarks by cliniciansboth imageshiddenunverifiedunknownleft out
its images or its truth are not released
Data availability statement not checked; likely on request.
3D multimodal dental CBCT and oral scanCBCT to intra-oral surface scan, same patient289nonedirectCC BY 4.0opentruth to make
no truth on both images; Two people click 6-10 cusp tips and incisal edges on the CBCT tooth surface and on the intra-oral scan of 10 patients (inter-observer error reported).
we would have to make the ground truthTwo people click 6-10 cusp tips and incisal edges on the CBCT tooth surface and on the intra-oral scan of 10 patients (inter-observer error reported).
4D laryngeal CT with inter-arytenoid distance (Monash)dynamic 4D laryngeal CT, frame to frame of the same subject, Parkinson's disease and controls12 subjects (8 Parkinson's, 4 controls), many frames eachmanual inter-arytenoid distance measurements on the framesboth imagesdirectCC BY 4.0opentruth to make
the truth is inter-arytenoid distances, not point pairs; landmarks would have to be made.
distances, not point pairs; made to validate an automatic measurement tooltwo people mark both arytenoid tips and 4 cartilage landmarks on 5 frames of each of 10 subjects; the measured distances serve as a check
75-Speaker Annot-16 (articulator contours)2D speech rtMRI of 16 speakers from the USC 75-speaker set, frames of the same video16 speakers, all frames with semi-automatic contours; 160 frames with handmade contourssemi-automatic articulator contours (tracked, expert-checked) on all frames; consensus handmade contours of 9 articulators on 160 framesboth imagesdirectCC BY 4.0openwaits for a tool
2D real-time MRI video.
Contours are vertex lists along articulators: near-point truth if vertex order is consistent (check)
AAPM RT-MAC 2019T2 MR simulation scans, head and neck; inter-subject55 (31 with contours)Manual parotids, submandibulars, nodal levels II/IIIboth imagescontrolledNIH Controlled Data Access Policyaskneeds a request
the data needs a request or a data use agreement
Status change: moved to controlled access (face policy).
ACRIN 6685-Tumor-Annotations (on ACRIN-HNSCC-FDG-PET-CT)FDG PET/CT, CT and some MR of head and neck cancer, same patient across trial timepoints (baseline, on-treatment, follow-up)250 patients annotated; 198 with >=2 annotated studies; 295 lesions tracked across studies (195 patients)NCI annotation project: up to 5 RECIST/PERCIST lesions per scan segmented, same lesions tracked over time (DICOM Tracking ID/UID) plus seed points; radiologists, US board-certified reviewboth imagescontrolledannotations CC BY 4.0; images NIH Controlled Data Access Policyaskneeds a request
the data needs a request or a data use agreement
Sparse lesion correspondences (centroids/seed points with tracking IDs) and lesion masks on both scans; lesions change size so use as landmarks, not dense overlap. Mostly pre-surgery PET/CT + CT pairs (multi-modal, near same date). Images moved to dbGaP controlled access (phs004225, head images) in July 2025.
ALI-CBCT landmarks (Michigan DCBIA)Large/medium field CBCT; inter-subject (orthodontic, some longitudinal in the group's AREG work)143 CBCT32 landmarks by 2 clinicians (119 in later models)both imageshiddenunverified (code open in Slicer)unknownleft out
its images or its truth are not released
Ask Cevidanes/Prieto lab for data; longitudinal orthodontic CBCT pairs would be valuable.
ARAR0331-Tumor-Annotations (on ARAR0331 (COG rhabdomyosarcoma))MR, CT, some PET of pediatric parameningeal rhabdomyosarcoma, same patient across trial timepoints (baseline, on-treatment, follow-up)108 patients; 103 with >=2 annotated studies; 277 lesions trackedNCI annotation project: up to 5 RECIST/PERCIST lesions per scan segmented, same lesions tracked over time (DICOM Tracking ID/UID) plus seed points; radiologists, US board-certified reviewboth imagescontrolledannotations CC BY 4.0; images NCTN/NCORP licence, now NIH Controlled Data Access Policyaskneeds a request
the data needs a request or a data use agreement
Sparse lesion correspondences (centroids/seed points with tracking IDs) and lesion masks on both scans; lesions change size so use as landmarks, not dense overlap. NCTN trial images: TCIA lists them as "Unavailable" since July 2025 (moving to NCI CTDC, dbGaP request, trial-specific PHS); annotations themselves are public.
Christchurch head and neck CT and CBCT with in silico deformations (Lowther 2020)planning CT and week 1 CBCT warped to later anatomy (in silico reference CBCTs), same patient12 patients x 3 later CBCTsground-truth DVFs from B-spline DIR of real later CBCTsboth imageshiddennot releasedaskleft out
its images or its truth are not released
Truth comes from another DIR, so it favours similar algorithms; not public.request; or same construction on public HNSCC-3DCT-RT or Head Neck CBCT CT (HUST)
CPTAC-HNSCC-Tumor-Annotations (on CPTAC-HNSCC)CT, PET/CT, few MR of head and neck cancer, same patient across trial timepoints (baseline, on-treatment, follow-up)99 patients annotated; 35 with >=2 annotated studies; 72 lesions trackedNCI annotation project: up to 5 RECIST/PERCIST lesions per scan segmented, same lesions tracked over time (DICOM Tracking ID/UID) plus seed points; radiologists, US board-certified reviewboth imagescontrolledannotations CC BY 4.0; images NIH Controlled Data Access Policyaskneeds a request
the data needs a request or a data use agreement
Sparse lesion correspondences (centroids/seed points with tracking IDs) and lesion masks on both scans; lesions change size so use as landmarks, not dense overlap. Images moved to dbGaP controlled access (phs004225, head images) in July 2025.
Crouzon-PUACT pre- and post-operative airway CTupper airway CT before and after midface advancement, Crouzon syndrome, same patient34 patients, 68 scansairway annotations (details unverified)requestno licence; a data use agreement (non-commercial, no redistribution, 24-month term)askneeds a request
restricted: a ScienceDB request, a signed data use agreement and the provider's approval; the agreement forbids uploading derived data and asks for deletion after 24 months
Large surgical change of bone and airway; Chinese Science Data Bank. Licence checked on 2026-10-10 at https://api.datacite.org/dois/10.57760/sciencedb.36271; derived test pairs shareable: unclear; commercial use: unclear. single 604 MB file; 2026 recordtwo people mark 15 skull points unchanged by surgery plus moved maxillary points
DentPair-850 (Peking University)CBCT tooth surface to intraoral scan mesh, same jaw, orthodontic patients850 jaw pairs from 425 patients4x4 transform per pair (how it was made: unverified)both imagesdirectCC BY-SA 4.0openwaits for a tool
tooth surface meshes from CBCT against intraoral scan meshes, not image pairs.
Meshes, not images; Chinese Science Data Bank. Complements the STSR CBCT-IOS task.two people mark crown cusps on 20 pairs to check the transforms
DWI and T2W head and neck MRI registration QA (MD Anderson)T2-weighted MRI to diffusion-weighted MRI, same session, pre-radiotherapy20 patientsmanual contours of parotids, submandibular glands, cord, brainstem and GTV on both sequences by a radiologist; 3 more observers on 5 casesboth imagesdirectCC BY 4.0openleft out
withdrawn by its owners from figshare between May and October 2026 (the page answers 410); no other copy
EPI distortion makes this a real intra-session multimodal deformable task.
French speakers real-time 2D and static 3D speech MRI (Isaieva 2021)2D real-time midsagittal MR at 50 fps plus 3D static phoneme volumes, 10 French speakers, frames and 2D-to-3D10 speakers x 16 series (~2,000 frames each) plus 76 static volumes eachphoneme and swallowing time stamps onlyone imagedirectCC0 (figshare items)opentruth to make
no truth on both images; articulator masks with the Ruthven 2D U-Net plus two people's tongue and velum points on 50 frames
Fails spatial truth; 3D static volumes at the same phonemes allow 2D-3D matchingarticulator masks with the Ruthven 2D U-Net plus two people's tongue and velum points on 50 frames
HaN-Seghead and neck CT and MR, same patient42manual contours on the CT onlyone imagedirectCC BY-NC-ND 4.0non-commercialleft outDraw 5-10 OARs (parotids, mandible, brainstem) on MR for 10 test patients: contours on both; two people add 15 bony landmarks on CT and MR.
Head and neck planning CT and recurrence CT (elastix, Strahlenther Onkol 2016)planning CT to recurrence CT (median 8.5 months later), HNSCC, same patient20 patientsmanual structures by one observer (details unverified)hiddennot releasedaskleft out
its images or its truth are not released
Long interval and anatomy loss; not public.request; bony and vascular landmarks by two people
Head-Neck Cetuximab (RTOG 0522)Head and neck cancer: RT planning CT plus pre- and post-treatment FDG PET/CT111 patientsRT structures on planning CT onlyone imagecontrolledNIH Controlled Data Access Policyaskneeds a request
the data needs a request or a data use agreement
Labels on one image only. Images moved to dbGaP controlled access (phs004225, head images) in July 2025.PET and CT of one PET/CT scan are aligned by the scanner, so a known warp on the CT gives exact PET-CT truth; for planning CT vs PET/CT, two people place 15 bony and vessel landmarks on 10 pairs, TotalSegmentator head-neck labels on both as a weaker check.
Head-Neck-PET-CTFDG-PET/CT to separate planning CT, same patient207 patients with a separate planning CT (298 in total)GTV contours drawn on planning CT and propagated to PET/CT by deformable registration (93 drawn directly on PET/CT)both imagescontrolledimages under NIH controlled data access; clinical data CC BY 3.0askneeds a request
the data needs a request or a data use agreement
Circular truth: PET/CT contours come from DIR. Listed under the PET/CT gap.
HECKTOR 2021-2025 (head and neck PET/CT)FDG-PET and low-dose CT from one PET/CT session, oropharyngeal cancer; subset with planning CT and RT dose (2025)~1,123 patients from 10 centres (2025); planning CT for a subset (287 with RT dose)GTVp and GTVn contours by experts on PET/CTboth imagesrequestdata by approved request on the challenge site; terms not shown publiclyaskneeds a request
the data needs a request or a data use agreement
Head and neck moves little during PET/CT, so alignment is close to exact. 2022 page says access paused while agreements are renewed; 2025 test set hidden (Docker). Planning CT vs diagnostic PET/CT is a real inter-session pair. Licence checked on 2026-10-10 at https://hecktor25.grand-challenge.org/participation-policies/; derived test pairs shareable: unclear; commercial use: unclear. policies only say participants may use the training data in any way for training modelswarp CT by a known field and register PET: exact truth; for planning CT vs PET/CT-CT, two people place 15-20 bony and soft-tissue landmarks on 10-20 pairs
HNSCC (incl. Head-Neck-CT-Atlas)Head and neck cancer: diagnostic CT/PET-CT/MR, RT planning CT, and follow-up imaging before and after RT627 patients; 215 (CT-Atlas) with whole-body PET-CT or abdominal CT before and after RTRT structures and dose on planning CT; no labels on follow-up scansone imagecontrolledNIH Controlled Data Access Policy (images); CC BY 3.0/4.0 (clinical)askneeds a request
the data needs a request or a data use agreement
Rich longitudinal multi-modal data but labels on one image only. Images moved to dbGaP controlled access (phs004225, head images) in July 2025.PET and CT of one PET/CT scan are aligned by the scanner, so a known warp on the CT gives exact PET-CT truth; for planning vs follow-up CT, two people place 15 bony and vessel landmarks on 10 pairs, TotalSegmentator labels on both as a weaker check.
HNSCC-3DCT-RT3 CT per patient during radiotherapy31 patients × 3contours (whether mid and post-treatment contours exist is unverified)controlledcontrolled access (dbGaP since 2025)askneeds a request
the data needs a request or a data use agreement
Used in 1 of the 2D and microscopy registration papers read (PolyPose 2025).Two people place 15 landmarks (vertebrae, mandible, hyoid, vessel branches) on all three CTs of 10 patients; TotalSegmentator head-neck labels on all three as a weaker check. Check whether mid and post contours exist.
Multiscale human inner ear dataset (Gerber 2017, SMIR)cadaver temporal bones, clinical CBCT and micro-CT, intact and after cochlear-implant electrode insertion, same specimen52 bones: 30 CBCT and 50 micro-CT volumes; implanted subsets (7 and 4 specimens) imaged after insertionmanual cochlea segmentations on 24 micro-CT; electrode, scala tympani and vestibuli segmented on implanted micro-CTone imageaccountnone stated in the Sci Data descriptor; data on the SICAS Medical Image Repository (smir.ch, returns 403 to scripts)unknownneeds an account
the download needs an account
Hosted on the SICAS Medical Image Repository (smir.ch); current availability unverified. Licence checked on 2026-10-10 at https://pmc.ncbi.nlm.nih.gov/articles/PMC5604133; derived test pairs shareable: unclear; commercial use: unclear. SMIR account needed; per-record SMIR licence not readableelectrode contacts bright on CBCT and micro-CT of implanted bones: contact centroids = landmarks across scales; intact vs implanted: two people place round window, apex and modiolus points; limits: metal blooming on CBCT, dummy arrays in one subset
NKI head and neck CT to CBCT with implanted gold markers (Mencarelli 2014)planning CT to weekly CBCT, oropharyngeal cancer, same patient13 patients, weekly CBCTs4 to 10 submucosal gold markers around the tumour (digitally removed before DIR) and 15 anatomical landmarks by two observersboth imageshiddennot releasedaskleft out
its images or its truth are not released
Rare soft-tissue marker truth in head and neck; not public.request from NKI-AVL
NPC planning CT and fractional CBCT (Li 2017)planning CT to CBCT, nasopharyngeal cancer, same patient21 patients, 129 CBCTsparotid, submandibular gland, vertebra, vertebral foramen drawn by one physician on CT and CBCT (repeated after three months)both imageshiddennot releasedaskleft out
its images or its truth are not released
Not public; no markers.request; bony landmarks by two people on CT and CBCT
OCTA-500OCT and OCTA volumes, 500 subjects, single visitNo longitudinal pairsVessel, FAZ and layer labelsrequestCC BY 4.0 (IEEE DataPort open-access record, JSON-LD licence); archive password obtained by an application e-mail to the authorsopenneeds a request
the data needs a request or a data use agreement
Fails pair criterion. Used in 1 of the 2D and microscopy registration papers read (Vanderbilt OCT mosaicking 2023 (synthetic mosaics)). Licence checked on 2026-10-10 at https://ieee-dataport.org/open-access/octa-500; derived test pairs shareable: yes; commercial use: yes. the password step is an access formality; licence itself is CC BY 4.0OCT and OCTA come from one acquisition, aligned by construction: a known warp on the OCTA projection vs the OCT projection gives exact multimodal truth. No longitudinal pairs.
OLIVESLongitudinal OCT volumes (49 B-scans) and fundus, DR/DME trial eyes; same eye over ~2 years96 eyes, 1,268 visitsPer-B-scan biomarker labels and clinical labels; no correspondencesone imagedirectCC BY 4.0opentruth to make
no truth on both images; Two people mark 10 vessel bifurcations on the fundus or OCT en-face image at two visits of 20 eyes
Fails truth for registration.Two people mark 10 vessel bifurcations on the fundus or OCT en-face image at two visits of 20 eyes; AutoMorph vessel junctions verified by a person speed it up.
OpenEarCBCT to micro-slicing of the temporal bone, same specimen and inter-specimen8 same specimen, 56 between specimensmanual labels registered onto the CBCTboth imagesdirect: Zenodo record 1473724; zips readable by ranges (the registered volumes and segmentation are about 0.4 GB a specimen)CC BY 4.0 (Zenodo and the Sci Data paper)openin the corpuspriority 2
Built: OpenEar (and OpenEar_Seg).
OpenKBPHead and neck planning CT resampled to 128^3 with OAR/target masks and dose; inter-subject340 patientsClinical contours from 12 institutionsboth imagesdirect: GitHub repository, provided-data/ (2.1 GB of CSV, a few MB a patient, files fetched one by one)MIT (repository LICENSE; data-description: CT and contours cleaned from TCIA, doses from synthetic plans)askin the corpuspriority 3
The repository (MIT) ships 340 head and neck planning CTs with the original clinical contours of seven organs at risk and the PTVs, all downsampled to 128 x 128 x 128 voxels of about 3.5 x 3.5 x 2 mm, as sparse CSV (data-description.pdf); an organ is missing where the clinic did not contour it. The CTs come from four TCIA collections, among them Head-Neck Cetuximab (RTOG 0522), the source of the corpus's PDDCA (OpenKBP paper refs 22-25); patient ids are anonymised (pt_1 to pt_340), so an overlap with PDDCA's cases cannot be excluded without comparing images. Head and neck CT between patients is already ranked on PDDCA, at full resolution.
Low resolution; parent TCIA collections now partly controlled. Licence checked on 2026-10-10 at https://github.com/ababier/open-kbp. data-description.pdf: images from TCIA, doses synthetic; keep MIT notice and TCIA citations
Paired head CT-MRI (120 neurosurgical patients)Thin-slice CT and MRI, same patient120 patientsNoneone imagedirectCC BY-NC 4.0non-commercialtruth to make
no truth on both images; Two people place 15 landmarks (skull, ventricles, vessels) on CT and MRI of 10 patients
Synthesis dataset; fails truthTwo people place 15 landmarks (skull, ventricles, vessels) on CT and MRI of 10 patients; SynthSeg labels on both as a weaker check.
Pedi-Cranial-CT-Healthy (TCIA 2026, FDA)non-contrast head CT of healthy children 1 month to 10 years, between subjects100 subjects (10 per age year)noneone imagedirectunverified (TCIA collection licence)unknowntruth to make
no truth on both images; two people place 15 cranial landmarks (nasion, sutures, foramina) on 20 subjects
Fails truth; useful for age-spanning atlas registration with large shape change.two people place 15 cranial landmarks (nasion, sutures, foramina) on 20 subjects; skull and brain masks from a public segmenter as a weaker check
Pre/post-orthodontic 3D dental models (Zenodo 11392406)Intraoral scan meshes before and after orthodontic treatment; same patient1,060 pre/post pairs from 435 patientsTooth segmentation, tooth position and crown landmarks on both scansboth imagesrequestCC0 (Zenodo), files restrictedopenneeds a request
the data needs a request or a data use agreement
Meshes, teeth move independently; landmark correspondence on both sides.
QIN-HEADNECKHead and neck cancer: FDG PET/CT before and after therapy, with follow-up scans279 patientsQIICR DICOM SEG and SR measurements on PET/CT (which timepoints are segmented: unverified)controlledNIH Controlled Data Access Policy (images); CC BY 3.0 (clinical)askneeds a request
the data needs a request or a data use agreement
Images moved to dbGaP controlled access (phs004225, head images) in July 2025.PET and CT of one PET/CT scan are aligned by the scanner, so a known warp on the CT gives exact PET-CT truth; across visits, two people place 15 bony and vessel landmarks on 10 pairs; QIICR segmentations, where on both time points, give a contour check.
RADCUREplanning CT, head and necksingle CT per patientcontoursone imagecontrolledcontrolled access (dbGaP since 2025)askneeds a request
the data needs a request or a data use agreement
One CT per patient: no pairs, only a known smooth warp (synthetic).
Real-time speech MRI with articulator segmentations (Ruthven 2023)2D real-time midsagittal MR of speech (10 fps), 5 volunteers, frames of the same series392 frames (67-105 per volunteer)manual masks of head, soft palate, jaw, tongue, vocal tract, tooth space on every frame (one MRI physicist; intra-rater repeat on 10%)both imagesdirectCC BY 4.0openwaits for a tool
2D real-time speech MRI.
Every-frame articulator labels: label truth for large non-rigid tongue motion; tinytwo people click tongue tip, velum tip, lip points on 50 frame pairs
SegRap2023Contrast and non-contrast H&N CT of same nasopharyngeal cancer patient120 train patients (200 total)45 OARs + 2 GTVs by oncologists (which CT carries labels unverified)accountSegRap2023 End-User Agreement (signed by a legal institution): scientific research only; no commercial use, no reproduction or distribution in any form, no sharing with third partiesnon-commercialneeds an account
the download needs an account
Same-session pairs; check whether pre-aligned Licence checked on 2026-10-10 at https://drive.google.com/file/d/1soBfTt78UBVL53SBBSXAvCVK0PkDbPI_/view; derived test pairs shareable: unclear; commercial use: no. agreement PDF linked from segrap2023.grand-challenge.org/dataset/; derived copies fall under 'reproduce ... in any form'Check the contrast and non-contrast CTs are not pre-aligned; two people place 15 bony landmarks on 10 pairs, and OARs give contours on both if both CTs carry labels.
STSR 2025/2026 Task 2: CBCT-IOS teeth registrationDental CBCT vs intraoral scan meshes (upper/lower jaw), same patient30 labelled + 300 unlabelled train pairs; 50 val pairs (hidden labels); 198 test pairsRigid 4x4 transform per jaw (organisers)requestResearch only, no redistribution or commercial use (dataset page)non-commercialneeds a request
the data needs a request or a data use agreement
Mesh-to-volume; organisers note left-handed coordinate inconsistencies in some casesOrganiser rigid transforms exist for the 30 labelled cases; two people click 6 cusp tips per jaw on 10 cases to check them, given the coordinate inconsistencies. Research-only licence.
SwallowReg tongue cine-MRI (MICCAI 2026)2D cine MRI of the tongue during swallowing, post-operative patients, frame to frame of the same patient (coronal and sagittal sets)moving/fixed frame pairs as .npz (count unverified)segmentation masks stored with each pair (structure and annotator unverified)both imagesdirectCC BY 4.0openwaits for a tool
2D cine MRI frames.
small preprocessed release; original cine series not included as far as the record showstwo people track 10 tongue-contour and hyoid points through 10 sequences; tagged MRI would give a physical reference
Teeth imaged by photoacoustic tomography, CBCT and micro-CT (Dryad 2022)in vitro extracted teeth imaged by PAT, CBCT and micro-CT, same toothcount unverifiednone; micro-CT can serve as referencedirectCC0opentruth to make
no truth on both images; cusp tips and root apices marked by two people on all three modalities of 10 teeth
rigid multimodal problemcusp tips and root apices marked by two people on all three modalities of 10 teeth
Teeth3DS+ / 3DTeethLandIntraoral scan meshes, upper/lower jaws; inter-subject (no images)1,800 scans (900 patients); 240 with landmarks publicExpert tooth labels and crown landmarks (mesial, distal, cusps, facial axis)both imagesdirectCC BY-NC-ND 4.0 (license.txt in the OSF Teeth3DS+ project; OSF licence field 'Other')non-commercialwaits for a tool
intraoral scan meshes, no images; ND licence.
Mesh registration only; no CBCT pairs. Licence checked on 2026-10-10 at https://osf.io/download/9dutn/; derived test pairs shareable: unclear; commercial use: no. website text is CC BY-SA 4.0 but the data licence file is BY-NC-ND; ND bars sharing adapted (resampled) pairs publicly
Temporal Bone Implants (Fairdata, Finland)six cadaver ears with cochlear-implant electrode arrays, high-resolution X-ray imaging6 ears; pre-insertion scans unverifiednone statedone imagerequestCC BY 4.0 (open preview record); the full PAS dataset (doi 10.23729/248dd82e-...) is 'Open' access with no licence listedopenneeds a request
the data needs a request or a data use agreement
Only a preview is public; data by email to fairdata-pas@gtk.fi. Modality (micro-CT likely) unverified. Licence checked on 2026-10-10 at https://metax.fairdata.fi/rest/v2/datasets/d16a6087-2ad5-48db-9d48-d678baaad6bd; derived test pairs shareable: yes; commercial use: yes. six ears with cochlear implant electrodes; check the full PAS record's terms before using it beyond the previewif a clinical CT or pre-insertion scan exists, electrode contact centroids give point truth across scans
ToothFairy (2023)Mandible CBCT; inter-subject443 CBCT, 153 with dense canal labelsManual inferior alveolar canal labels (dense on 153, sparse on rest)both imagesaccountCC BY-NC-SA (unverified for v1)non-commercialneeds an account
the download needs an account
Single structure; superseded by ToothFairy2/3.
ToothFairy2 / ToothFairy3Maxillofacial CBCT, 2-3 scanners; inter-subject532 public CBCT (ToothFairy3), 480 in ToothFairy2Manual labels: 42 classes (TF2), 77 classes incl. each tooth, pulps, canals, sinus, pharynx (TF3)both imagesaccountCC BY-NC-SA 4.0 (ditto page)non-commercialneeds an account
the download needs an account
Best dental inter-subject label set; no repeat scans.
UCL head and neck CT to CBCT before replanning (Veiga 2014)planning CT to mid-treatment CBCT and replan CT, same patient5 patients (conference paper); Med Phys 2014 cohort size unverifiedC1, C4, C7, body contour and sternocleidomastoid muscles drawn by one expert on CT and CBCT; replan CT dose as referenceboth imageshiddennot releasedaskleft out
its images or its truth are not released
NiftyReg evaluation paper; not public.request from UCL
USC 75-speaker speech production rtMRI (Lim 2021)2D midsagittal real-time MR of speech at 83 fps plus 3D vocal-tract volumes and T2w anatomy, 75 speakers, frames and 2D-to-3D75 speakers, ~17 min of video eachnone in the base release (audio aligned)one imagedirectCC BY 4.0opentruth to make
no truth on both images; articulator contours from Annot-16
Raw k-space included; Annot-16 adds contours for 16 speakersarticulator contours from Annot-16; tongue tip and lip points by two people
USC-TIMIT2D real-time midsagittal speech MR of 10 speakers plus electromagnetic articulography (EMA) of 4 speakers on the same 460 sentences10 speakers x 460 sentencesnone in the images; EMA sensor tracks are from a separate sessionone imagedirectno licence: 'The database, and companion software tools, are freely available to the research community'non-commercialtruth to make
no truth on both images; articulator contours by a public segmenter
EMA is not simultaneous with MRI, so no direct point truth Licence checked on 2026-10-10 at https://sail.usc.edu/span/usc-timit/; derived test pairs shareable: unclear; commercial use: unclear. research-community wording only; commercial use not addressedarticulator contours by a public segmenter; two people click tongue tip and lips on 50 frame pairs
Virtual head and neck phantom benchmark (JACMP 2021)virtual H&N CT phantoms with anatomy changecount unverifiedknown deformation vector fields and contoursboth imagesrequestnone stated: kVCT SOT/sEOT downloadable from the DIREP website; full set with contours and DVFs 'available through Oncology Systems Limited' (commercial vendor)askneeds a request
the data needs a request or a data use agreement
Data from the corresponding author only. Licence checked on 2026-10-10 at https://pmc.ncbi.nlm.nih.gov/articles/PMC8130232; derived test pairs shareable: unclear; commercial use: unclear. DIREP site terms not checked (no URL in the paper); ground-truth DVFs come from a commercial product
Weekly intra-treatment DWI on MR-Linac, head and neck (MD Anderson)weekly DWI and ADC during MR-Linac radiotherapy, same patientpatients and weeks unverifiedtumour ROIs per week (unverified)figshare, direct, one 239 MB zip readable by rangesCC BY 4.0 (figshare)openin the corpuspriority 2
30 HNSCC patients treated on the 1.5 T Unity MR-Linac, ADC maps at baseline, weekly (W1 to W6) and after RT (4 to 8 sessions a patient), each with an RTSTRUCT: the baseline GTVp on the baseline images, GTV-N and the residual primary (GTVP-RD) drawn on every weekly MRI, reviewed by two radiation oncologists; the weekly sets also hold GTV_P_BL, and the provider aligned the weeks to baseline with Velocity's deformable registration (zip listing, Sci Data 2024). As in HNTSMRG, the truth is one or two tumour centres a pair on volumes that shrink, so ranked apart; only the ADC is released (no T2w), but under CC BY 4.0 where HNTSMRG is non-commercial.
Longitudinal MR-linac; check which masks ship.one expert contours GTV and parotids weekly, a second on 5 patients

Spine and MSK · 25

DatasetProblemPairsChecked againstAccessLicenceStatusNotesTo make the truth
A091105-Tumor-Annotations (on A091105 (desmoid tumours, sorafenib))MR and CT of desmoid tumours (trunk, limbs, abdomen), same patient across trial timepoints (baseline, on-treatment, follow-up)83 patients; 82 with >=2 annotated studies (978 studies); 101 lesions trackedNCI annotation project: up to 5 RECIST/PERCIST lesions per scan segmented, same lesions tracked over time (DICOM Tracking ID/UID) plus seed points; radiologists, US board-certified reviewboth imagescontrolledannotations CC BY 4.0; images NCTN/NCORP licence, now NIH Controlled Data Access Policyaskneeds a request
the data needs a request or a data use agreement
Sparse lesion correspondences (centroids/seed points with tracking IDs) and lesion masks on both scans; lesions change size so use as landmarks, not dense overlap. NCTN trial images: TCIA lists them as "Unavailable" since July 2025 (moving to NCI CTDC, dbGaP request, trial-specific PHS); annotations themselves are public.
AIHub degenerative spine multimodality data (Korea, dataset 611)spine images from 8 hospitals, several modalities (CT, MRI, X-ray stated), same patientover 1.39 million annotated images; per-patient multimodal pairing unverifiedannotations of 4 degenerative diseases (type unverified)controlledAIHub terms; online safe zone, Korean nationals onlyaskneeds a request
the data needs a request or a data use agreement
found on AIHubvertebral body corners and pedicle centres on CT and MRI of 10 patients by two people
Brown carpal database (wrist anatomy and kinematics)wrist CT in many positions, same subject; released as bone surfaces plus kinematics90 subjects, 120 wrists, 1,215 positionsper-bone rigid kinematics, segmented carpal, radius and ulna surfacesboth imagesaccountBrownUniversity Carpal Dataset, MIT Use Agreementopenneeds an account
the download needs an account
Fails the image criterion: CT images not included, only surfaces.
DeepUS-ReconSeg (multi-angle forearm B-mode ultrasound)B-mode ultrasound frames of forearms in several orientations, same subject14 subjects, 4,200 frames (100 per orientation)noneone imagedirectunverified (Mendeley Data)unknowntruth to make
no truth on both images; a tracked probe would give poses
Frame-to-frame or view-to-view; fails truth.a tracked probe would give poses; otherwise two people mark 8 tendon/bone landmarks on 20 frame pairs
Four-dimensional CT of the knee for MPFL planning (Science Data Bank)dynamic 4D CT of active knee flexion-extension, same knee66 knees, 20-30 frames eachlandmark tables and kinematics (controlled access)controlledCC BY-NC 4.0non-commercialneeds a request
the data needs a request or a data use agreement
Open part is point clouds only; DICOM on request.bones are rigid: rigid bone poses per frame as truth
In vivo validation dataset for dynamic volumetric MRI (UMC Utrecht, thigh under pressure cuff)3D MRI of the thigh deformed by a pressure cuff, same subject; undersampled dynamic k-space plus fully sampled static scans at nine pressure levels9 volunteers x 9 pressure levels (36 pairs per volunteer if all levels are paired); 4 dynamic scans eachfully sampled reference images at each pressure level; muscle segmentation on the anatomical scan onlyone imagedirectCC BY 4.0opentruth to make
no truth on both images; muscle labels propagated to each pressure level and corrected by a person
From the MR-MOTUS group (MR-guided radiotherapy). Raw k-space (MRD/HDF5), images must be reconstructed with BART. Deformation truth between pressure levels is not given.muscle labels propagated to each pressure level and corrected by a person; or 15 vessel and fascia landmarks per level by two people
IWOAI 2019knee DESS MR baseline to 1 year, same patient88cartilage and menisci by one expertboth imagesrequestrepo AGPL-3.0 covers code only; images are OAI data (NIMH Data Archive terms), segmentations distributed by the challenge organisers on requestaskneeds a request
the data needs a request or a data use agreement
Licence checked on 2026-10-10 at https://github.com/denizlab/2019_IWOAI_Challenge; derived test pairs shareable: no; commercial use: unclear. OAI DUA applies to the images; organiser terms for the labels not public
KMAR-50K (paired knee MRI with motion artefact and rescan)knee MRI with motion artefact and the rescan of the same sequence, multi-view, same patient1,444 sequence pairs from 1,190 patientsnone independent; rescans also given registered to the artefact scan with ANTs SyNone imagedirectunverified (Mendeley Data)unknowntruth to make
no truth on both images; two people place 10 bone landmarks on 20 pairs
Sci Data 2025. Two Mendeley repositories (xw7mrg7ntg and 95w9f5tzz8).two people place 10 bone landmarks on 20 pairs; bone masks from a public segmenter on both scans
LumbarSR (clinical CT and photon-counting micro-CT)clinical CT in 8 settings to photon-counting micro-CT, dry lumbar vertebrae30 specimens x 8 settingsrigid ANTs alignment to micro-CT; bone masksboth imagesdirectCC BY 4.0opentruth to make
the truth is a program's rigid alignment (ANTs) of dry vertebrae, circular.
Truth is a program's rigid alignment; dry bone only.two people mark 10 cortical landmarks per vertebra
MMCSD (cervical spondylosis CT and MRI)cervical spine CT (bone and soft tissue windows) and sagittal T1/T2 plus axial T2 MRI, same patient250 patientsnoneone imageaccountunverified (Synapse syn63903115)unknownneeds an account
the download needs an account
Sci Data 2025.vertebra masks from a public segmenter on CT and MRI (weaker); two people place 2 landmarks per vertebra on 10 patients
MOST (Multicenter Osteoarthritis Study)knee radiographs, full-limb radiographs and knee MRI, same knee from baseline to 168 months3,026 original + 1,500 new participants; visits at 0, 15, 30, 60, 84 and 144-168 monthssemi-quantitative OA readings (no spatial truth)requestcontrolled access through an NIA AgingResearchBiobank request; study record: 'Materials from this study are not available for commercial use'; data 'some restrictions apply'non-commercialneeds a request
the data needs a request or a data use agreement
OAI's sister cohort (raw OAI listed). Licence checked on 2026-10-10 at https://agingresearchbiobank.nia.nih.gov/studies/most/details; derived test pairs shareable: unclear; commercial use: no. request with a sample agreement; images included in the biobank materialsa public knee segmenter on both (weaker); two people place 10 bone landmarks on 20 pairs
Multimodal 3D US and CT of vertebrae (spinal surgery database)3D ultrasound to CT of vertebrae: human lumbar (simulated US), canine and lamb vertebra phantoms (real and simulated US)3 datasets (1 human simulated, 2 ex vivo animal phantoms)21 fiducial landmark pairs per phantom dataset (authors)both imagesZenodo, direct; one zip holding three inner zips (CaninePhantom, HumanSubjects, Lamb)CC BY 4.0 (Zenodo)openin the corpuspriority 3
Built: SpineUSCT (two phantom pairs).
Tiny, ex vivo; the human set uses simulated US.
Osteoarthritis Initiative knee MRI, ICON/GradICON registration splitDESS knee MRI, between subjects (left/right flipped)2,532 train images, 301 test pairsfemoral and tibial cartilage segmentations (automatic, from the OAI processing of the Niethammer group; unverified)both imagesrequestOAI data under the NIMH Data Archive data use certification (collection 2343) per screening; NDA pages are JS-rendered, terms not re-readaskneeds a request
the data needs a request or a data use agreement
Distinct from the 'UK Biobank, NAKO, raw OAI' entry: a fixed registration split reused by the UNC group. The raw OAI needs the NDA agreement. Licence checked on 2026-10-10 at https://nda.nih.gov/oai; derived test pairs shareable: no; commercial use: unclear. uniGradICON lists the OAI split but does not redistribute images; NDA access needs an approved accounttwo people place 15 bone landmarks (tibial spines, condyles) on 20 test pairs; OAI-ZIB manual labels on both where the subjects overlap
QIN-SARCOMASoft-tissue sarcoma DCE-MRI at 3 visits during preoperative chemoradiation15 patients, 38 studiesNoneone imagedirectCC BY 4.0opentruth to make
no truth on both images; Two people place 10 landmarks (bone points, vessel branches) on 10 visit pairs, away from the tumour
Fails truth.Two people place 10 landmarks (bone points, vessel branches) on 10 visit pairs, away from the tumour; TotalSegmentator MRI bone and muscle labels on both as a weaker check.
R-D-Reg (cervical spine CT-MRI)Cervical spine CT vs MRI, same patientUnverified countAnnotations per paper (vertebrae; unverified)directno data licence; repo MIT covers code; dataset on a Quark cloud-drive linkunknownneeds a request
the Quark share link is blocked (Quark flags the file as violating its rules) and the README asks to email the authors for the dataset
Data via Quark cloud link Licence checked on 2026-10-10 at https://github.com/ssc1230609-spec/MSR-registration; derived test pairs shareable: unclear; commercial use: unclear. README 'Data' section gives only the download linkTwo people place vertebra centroids and endplate corners on CT and MRI of 10 pairs (piecewise rigid truth); TotalSegmentator vertebrae on CT and MRI as a weaker check.
Robotic and handheld lumbar spine US with CT (KU Leuven)tracked 2D ultrasound (handheld and robotic) to ultra-low-dose CT of the lumbar spine, same subject63 healthy volunteers; 223 handheld + 375 robotic US scans, each with a CTtracked probe or robot poses, CT surface models of L1-L5, 6,091 US frames with bone-surface labels (authors)both imagesdirectCC BY 4.0openwaits for a tool
tracked 2D ultrasound frames to CT (2D/3D).
Whether a reference US-to-CT transform per scan is included is unverified; bone labels exist on both sides.
Sheffield lower-limb muscle MRI (atlas registration inputs)Muscle MRI, inter-subject20 combinations of 5 subjectsFixed-image muscle segmentationsone imagedirectCC0opentruth to make
no truth on both images; Each of the 5 subjects is the fixed image in some combination, so its muscle labels exist: reuse them on the moving side for contours on both images of every pa
Only fixed labels in each fileEach of the 5 subjects is the fixed image in some combination, so its muscle labels exist: reuse them on the moving side for contours on both images of every pair.
Shoulder 3D ultrasound mosaickingRobot-tracked 3D US volumes + MRI bone models, same volunteer5 volunteers, 321 volumesCalibrated robot poses; refined registrations; humerus labels on 62 US volumes; MRI bone STLsdirect: Zenodo, one 8.5 GB zip whose members read by HTTP rangesCC BY 4.0 on the Zenodo recordopento build apartpriority 3
5 volunteers, 7 trials, 321 robot-tracked 3D US volumes (512x403x256 uint8 at 0.11 x 0.11 x 0.24 mm), humerus labels on 62 of them, the MRI only as humerus, scapula and clavicle STL meshes (no MRI image). The provider's truth between overlapping volumes is rigid (calibrated robot poses, ImFusion-refined registrations) and the bone is rigid, so labels leave no room after alignment: the only truth is a known warp on robot-aligned overlapping volumes (synthetic), and five subjects cannot give ten test and ten development pairs.
Calibrated robot poses give US-to-US truth; for US to MRI, the humerus labels on 62 volumes against the MRI bone model give a surface-distance check.
SKM-TEA (Stanford knee qDESS)knee qDESS MRI, echo 1 and echo 2 from one acquisition, raw k-space and DICOM155 scans (unverified)manual cartilage and meniscus segmentations; pathology boxesboth imagesaccountStanford AIMI shared-dataset licence (Stanford School of Medicine terms, as for the LERA dataset): free, restricted to non-commercial purposes; commercial entities can license AIMI datasets for a feenon-commercialneeds an account
the download needs an account
The two qDESS echoes are read in one sequence: aligned by construction, different contrast. Segmentations apply to both echoes. Licence checked on 2026-10-10 at https://aimi.stanford.edu/commercial-use-aimi-datasets; derived test pairs shareable: unclear; commercial use: no. SKM-TEA paper A.5 names the Stanford licence; download via the AIMI portal (account); repo code MITapply a known smooth warp to one contrast and register the other to it: exact multimodal truth (echo 2 to echo 1); warp the labels too for a Dice check
Soft-tissue-SarcomaExtremity soft-tissue sarcoma: T1 and T2-FS MRI and FDG-PET/CT of the same patient (MR/PET/CT)51 patients, one MR study + one PET/CT study eachTumour contours (GTV mass, GTV with oedema) drawn on T2-FS, propagated to PET and T1 by MIM rigid registration; RTSTRUCT on all scansboth imagesdirectCC BY 3.0opentruth to make
the labels were copied between images by the creators' rigid registration, circular.
Labels on all images but copied by registration: circular.
Spinal-Multiple-Myeloma-SEGDual-energy low-dose CT of the spine in multiple myeloma; a few patients scanned twice67 patients, 72 scans (about 5 repeat pairs)Manual vertebra (with level) and myeloma lesion masks on each scanboth imagesdirect: TCIA NBIA API, no account (masks also as NIfTI through a public Aspera link)CC BY 4.0 (collection page and every NBIA series)openin the corpuspriority 3
In the corpus; only 5 patients have repeat scans.
Reconstructions from one spectral acquisition need no registration; only the ~5 repeat scans give pairs. Published 2026-03.
Spine CT and 3D US (Masoumi et al.)Vertebrae CT vs real/simulated US: 3 patients (sim US), canine and lamb phantoms3 datasets21 CT-US landmark pairs on each phantom; fiducial silver standardboth imagesdirect: Zenodo, one 0.15 GB zipCC BY 4.0 on the Zenodo recordopensame as
Same Zenodo record (4813508, Data.zip, 0.15 GB, CC BY 4.0) as registration 462 and synthesis 150: three patients' lumbar CT with US simulated from it, and a canine and a lamb vertebra phantom with CT, real US, simulated US and 21 CT-US landmark pairs each (a fiducial silver standard).
Spine-Mets-CT-SEGSpine metastases RT simulation CT ("pre and post radiotherapy" in title)None: NBIA lists one study per patient (55)Vertebra level labels and lesion classone imagedirectCC BY 4.0opentruth to make
no truth on both images; One CT per patient: no pairs, no usable truth.
Fails pairs despite the title.One CT per patient: no pairs, no usable truth.
UC Davis total-body PET/CT of myofascial pain, FDG and butanol (Dryad)two total-body PET/CT scans per subject ([11C]butanol and [18F]FDG) on uEXPLORER, cropped to the lumbosacral regionone CT-CT and PET-PET pair per subject (subject count unverified)L4-S1 segmentation on each scanboth imagesDryad, no account, but files only through a browser (API download needs a bearer token, file_stream 403 to scripts)CC0 1.0 (Dryad)openleft out
The Dryad record (CC0, 8.84 GB in one zip plus README and two csv) gives neither the subject count nor how the images were labelled, and its files cannot be fetched by a script (the API asks for a token, the web file stream answers 403 behind a bot check), so the README stayed unread. The study's 2025 JNM abstract reports 15 participants imaged with a sequential dual-tracer protocol ([11C]butanol then [18F]FDG, apparently one session) and paraspinal (L4-S1) and gluteal muscles segmented on CT, method not stated: the two PET scans share one position, and fewer than twenty subjects. Worth one look at the README by hand before closing it.
same or separate session unverified; low-dose AC CT onlyvertebral landmarks (endplate corners, spinous tips) on both CTs by a program, verified by a person
USC Hand MRI datasethand MRI in 12 poses per subject, same subject multi-pose4 subjects x 12 poses = 48 scansnone listeddirectCC BY 4.0opentruth to make
no truth on both images; Two people place 15 landmarks (bone tips, joint centres) on 10 pose pairs
Images only (labels unverified).Two people place 15 landmarks (bone tips, joint centres) on 10 pose pairs; bones are rigid, so per-bone rigid fits give piecewise truth.

Vascular · 9

DatasetProblemPairsChecked againstAccessLicenceStatusNotesTo make the truth
4D-CTA aortic aneurysm kinematics (Perth, 10 patients)ECG-gated 4D CTA of abdominal aortic aneurysm over the cardiac cycle10 patients x 10 phasesone aneurysm surface model per patientone imagedirectCC BY 4.0opentruth to make
no truth on both images; segment the wall on every phase
Phase-to-phase motion is small.segment the wall on every phase; calcifications as landmarks
4D-CTA aortic aneurysm structural integrity (Perth, 20 patients)ECG-gated 4D CTA of abdominal aortic aneurysm20 patientsgeometries and biomechanics per patient (per-phase labels unverified)directCC BY 4.0opentruth to make
4D CTA with 2 to 10 phases a patient (NRRD) but one wall and thrombus geometry per patient (VTP), not one per phase
Extends the 10-patient set.segment the wall on every phase; calcifications as landmarks
AIHub carotid CT and MRI/MRA (Korea, dataset 71798)carotid MRI/MRA and CT of the same patientunverifiedunverifiedcontrolledAIHub terms; online safe zone, Korean nationals onlyaskneeds a request
the data needs a request or a data use agreement
found on AIHubcarotid bifurcation, vessel centrelines and calcifications marked on both modalities by two people on 10 patients
ImageTBAD (type-B aortic dissection CTA)aortic CTA, single time pointnone (100 CTAs)true lumen, false lumen and thrombus masksone imagedirectCC BY-NC 4.0 (Kaggle, uploaded by the author Xiaowei Xu)non-commercialtruth to make
no truth on both images; No follow-up scans: no pair truth
Fails: no follow-up scans; access unverified. Licence checked on 2026-10-10 at https://www.kaggle.com/datasets/xiaoweixumedicalai/imagetbad; derived test pairs shareable: yes; commercial use: no. Kaggle API licenseName; arXiv paper says dataset and code released to the publicNo follow-up scans: no pair truth; inter-subject lumen masks only, a weak check.
IntrA (intracranial aneurysm meshes)intracranial vessel and aneurysm surface meshes from MRA, single time pointnoneaneurysm segment labels on meshesone imagedirectnone stated; 'open-access' dataset on Google Drive, optional user formunknowntruth to make
no truth on both images; Meshes only, no image pairs: no usable truth.
Fails: no pairs, meshes only. status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Point-cloud shape survey 2025). Licence checked on 2026-10-10 at https://github.com/intra3d2019/IntrA; derived test pairs shareable: unclear; commercial use: unclear. no LICENSE file; README asks to citeMeshes only, no image pairs: no usable truth.
IVUS with EM tracking to CT (Dillon et al.)intravascular ultrasound reconstructed in 3D by EM tracking vs preoperative CT vessel7 phantoms, 3 animalsCT vessel meshes and centerlines registered to the IVUS/EM reconstruction; calibration transformsboth imagesdirectCC BY 4.0opentruth to make
the truth is the authors' registration of CT meshes to the IVUS reconstruction, circular; phantoms and 3 animals.
Labelled IVUS companion set: Zenodo 20752361.fiducials on the phantoms seen in CT and IVUS for a TRE
Longitudinal AAA CTA (Seoul National University and Amsterdam UMC)abdominal aortic aneurysm CTA follow-up, same subject138 CTAs from 31 patients, at least 3 scans eachlumen and thrombus segmentations (authors)both imageshiddenunknownleft out
its images or its truth are not released
No public release found; only longitudinal aorta set seen.
Repositionable 3D ultrasound (CMUT) microvascular monitoring (Dryad 2025)3D ultrasound of microvessels with a repositionable probe at several times, same sitecount unverifiednonedirectCC0opentruth to make
no truth on both images; vessel bifurcations found by a program and checked by a person across sessions
vessel bifurcations found by a program and checked by a person across sessions
TopAneu 2026 (intracranial aneurysm CTA and MRA)brain CTA and TOF-MRA with aneurysms; mostly one scan per patient, Geneva centre has both modalities, 7 patients with repeat CTA415 training scans from 408 patients, so very few intra-patient pairsexpert aneurysm masks with 50+ vessel-location classes; silver vessel masks from the TopBrain modelboth imagesaccountunverifiedunknownneeds an account
the download needs an account
Fails the pair criterion except the few CTA-MRA and repeat scans; usable for inter-subject vessel-label registration. Test set private.vessel bifurcations found by a program on the vessel masks then verified by a person, on the CTA-MRA and repeat pairs

Preclinical · 44

DatasetProblemPairsChecked againstAccessLicenceStatusNotesTo make the truth
Allen CCFv3, DevCCF, mouse lemur atlasa single template per age–no paired truthone imagedirectmixed: Allen CCFv3 under Allen Institute Terms of Use (research or other non-commercial use; a commercial service for third parties counts as commercial); DevCCF CC BY 4.0 (figshare); mouse lemur atlas not checkednon-commercialtruth to make
no truth on both images; Templates only: targets for atlas registration
Used in 4 of the 2D and microscopy registration papers read (STalign 2023; mBrainAligner 2022; Princeton Mouse Atlas 2021; NeuReg 2024 (DevCCF)). Licence checked on 2026-10-10 at https://alleninstitute.org/terms-of-use/; derived test pairs shareable: unclear; commercial use: no. DevCCF alone is open (figshare 25714965, 26377171, CC BY 4.0)Templates only: targets for atlas registration; truth must come from the subject images.
aMAP validation set (GIN brainglobe/amap_validation)STPT mouse brain stacks to Allen atlas (Kim 2014 atlas); between subjects6 brains downsampled to 12.5 um (2 parameter-search, 4 scoring)Manual segmentations of 9 structures by 22 neuroscientists, STAPLE agreementone imagedirectCC0 1.0opentruth to make
no truth on both images; Add 20 point landmarks per brain by two raters to complement Dice.
STPT not light-sheet; labels on 40-plane stacks only.Add 20 point landmarks per brain by two raters to complement Dice.
Animal Brain Collection (cross-species brain MRI and histology)postmortem brain MRI (T1, T2, DTI) and histology of the same brains, many vertebrate species44 MRI datasets plus histologynone for alignmentone imagedirectunverified (OpenNeuro and SSBD)unknowntruth to make
no truth on both images; two people place 10 landmarks per histology section on 10 brains
Sci Data 2025.two people place 10 landmarks per histology section on 10 brains
BigMac macaqueOne macaque: in vivo MRI, post-mortem MRI/dMRI, whole-brain PLI and myelin/Nissl histology1 brain, hundreds of slides to MRIAuthors' 2D-to-3D warpfields (their registration)one imagedirecthosted on the Oxford Digital Brain Bank ('openly available'); DBB says many datasets carry a material transfer agreement limiting use to research/education; BigMac's own terms not readable (JS app)asktruth to make
no truth on both images; Two people place 10 landmarks (vessels, sulcal fundi) on 10 histology or PLI section/MRI slice pairs, and 15 on in vivo vs post-mortem MRI
Large volume; no independent truth. Licence checked on 2026-10-10 at https://pages.fmrib.ox.ac.uk/amyh/bigmacdocumentation/data_download.html; derived test pairs shareable: unclear; commercial use: unclear. DBB paper (eLife 2022, PMC9042233) describes the MTA schemeTwo people place 10 landmarks (vessels, sulcal fundi) on 10 histology or PLI section/MRI slice pairs, and 15 on in vivo vs post-mortem MRI; the authors' warpfields are a reference only.
Brain/MINDS 3D marmoset atlas BMA-2017 (Woodward 2018)One marmoset: ex vivo T2 MRI vs reconstructed Nissl histology1 specimen200+ manual region labels per hemisphere; Nissl already registered to MRIboth imagesdirect: datasets.brainminds.jp/brainminds/BMA-2017, no accountCC BY 4.0 (LICENSE file in the repository)openleft out
One marmoset brain: an ex vivo T2 MRI, the reconstructed Nissl volume already registered to the shape of the MRI, the region segmentation and the gray/white and mid-cortical boundaries (README of the BMA-2017 repository, CC BY 4.0). One pair already aligned gives one synthetic case at most, and its region labels are the atlas that NA216 carries to every animal.
Raw (unregistered) sections availability unverified. Licence checked on 2026-10-10 at https://datasets.brainminds.jp/brainminds/BMA-2017/raw/branch/master/LICENSE; derived test pairs shareable: yes; commercial use: yes. Brain/MINDS data portal repository, initial release 2017-11-22
Brain/MINDS marmoset NA216 / eNA91Marmoset in vivo T1/T2/DTI/rsfMRI (216) and ex vivo high-res MRI (91 of them); in vivo vs ex vivo same animal, and inter-subject91 in vivo-ex vivo pairs; many inter-subjectPer-subject label maps (method not stated; likely atlas-propagated)both imagesdirect: datasets.brainminds.jp/brainminds/MRI-NA216 (GIN, git-annex files served by /media/ URLs, range requests), no accountCC BY 4.0 (portal page and repository LICENSE)openin the corpuspriority 2
File names link 62 of the 91 ex vivo T2 scans (9.4 T, 0.1 mm, 300 x 360 x 200) to the same animal's in vivo T2 (0.27 mm), and 30 of them also have ex vivo 52-region label maps next to in vivo ones: real in vivo to ex vivo pairs (fixation, extraction). The Sci Data paper says the labels are the Hashikawa histology atlas carried to each animal by an ANTs registration from the standard brain, so the truth is automatic and leans towards ANTs. CC BY 4.0, files downloadable without an account from datasets.brainminds.jp (GIN).
Truth probably automatic (unverified); in vivo-ex vivo pairs are rare and useful.
CUBIC-Atlas, CUBIC-Cloud and SSBD CUBIC brainsPoint-cloud single-cell atlas and immunostained whole brains mapped to it; SSBD has CUBIC LSFM brains (Susaki 2014); between subjectsCUBIC-Cloud: 9 stain sets; SSBD project 48: 27 image datasetsNone; CUBIC-Atlas annotated from ABA by registrationone imagedirectCUBIC-Atlas CC BY-NC-ND 4.0; SSBD CC BYnon-commercialtruth to make
no truth on both images; Landmarks (20) by two raters on SSBD brains at 25 um
Fails truth; mostly point clouds, not image pairs. ND licence blocks derived benchmarks for the atlas.Landmarks (20) by two raters on SSBD brains at 25 um; point-cloud to point-cloud check with cell counts per region (weak).
D-LMBmap LSFM registration dataiDISCO+/Adipo-Clear LSFM mouse brains (autofluorescence) to Allen CCFv3; between subjects12 brains with 6 region labels; 9 brains with 5 small-structure labels; 18 landmarks in atlas and LSFM brainsManual labels and 18 manual landmarks by the authors (whether files are in the source-data zips is unverified)both imagesdirect: Zenodo record 8123585 (three zips readable by ranges) and a public Google Drive folder (no account)CC BY 4.0 (Zenodo; the GitHub README says CC BY 4.0 too)openleft outpriority 3
Only one released landmark-annotated subject is usable for this task. Retained as an example outside the benchmark in datasets/examples/DLMBmapLSFM.
Check the zips for the landmark and label files before use.
DePoLL (Deformable Porcine Laparoscopic Liver)pre-operative CT liver model to intra-operative CT and laparoscopic video, in vivo pig, 13 deformed states13 intra-operative configurations vs 1 referenceabout 60 metal fiducials (surface clips, internal spheres) localised in intra-operative CTboth imagesdirectCC BY-NC-ND 4.0non-commercialto buildpriority 2
Pig liver under laparoscopic deformation with about 60 metal fiducials in 14 CTs fills a gap, but comes only as one 110 GB archive (videos included; the laptop has 143 GB free); ND licence.
Also usable as CT-to-CT (pre- vs intra-operative).
Ex vivo porcine bladders with fiducial markers (Wognum 2014)CT of excised porcine bladders filled from 100 to 400 ml, same organ5 bladders x 6 fillings against the 100 ml scan (about 30)30 to 40 radiopaque markers on the outer wall per bladder; bladder contoursboth imageshiddennot releasedaskleft out
its images or its truth are not released
Large controlled deformation with marker truth; markers are visible, so they should be erased before blind use.request from AMC Amsterdam; erase markers digitally and keep their positions as truth
Gubra LSFM mouse brain atlas and multimodal MRI/LSFM atlasiDISCO+ LSFM template vs Allen CCFv3 template; LSFM brains to template; MRI-LSFM bidirectional atlas (Perens 2023)Template pair (LSFM vs CCFv3); 1 sample brain (3 channels) released; 10 evaluation brains not released27 manual landmarks on the LSFM template (coordinates CSV + label image) and on CCFv3; per-brain landmark distances (CSV) for 10 brainsboth imagesdirect (GitHub repository, NIfTI files of tens of MB)CC BY 4.0 (repository LICENCE.txt)openleft out
The repository's landmarks are 27 points on the Gubra LSFM template only (gubra_template_wo_olf_landmark_coordinates.csv); for the ten evaluation brains (ID017 to ID041) it gives only per-landmark distances, not the brains or their points. The one sample brain it releases comes already aligned to the template (LSFM_atlas_aligned_autofluorescence), and CCFv3 is not in the repository and is under the Allen terms. No pair carries a truth; the only route is to ask Gubra for the ten brains.
Landmarks only usable template-to-template unless Gubra shares the 10 brains. Multimodal MRI/LSFM templates hosted elsewhere (unverified). Used in 1 of the 2D and microscopy registration papers read (Perens 2021).Ask for the 10 evaluation brains; place the same 27 landmarks on the released sample brain to make one real pair.
IMPC mouse embryo micro-CT (E14.5-E18.5) and LAMA E14.5 atlasIodine-stained micro-CT of wild-type and knockout mouse embryos at ~13-14 µm; between subjects to a population averageHundreds of embryos (>750 reconstructions in 2017); many WT-to-average pairsLAMA E14.5 population average with expert atlas labels (Zenodo 4559800); per-embryo labels: noneone imagedirectCC BY 4.0 (IMPC site: 'Content on this site is licensed under a Creative Commons Attribution 4.0 International license'; LAMA E14.5 atlas Zenodo 4559800 CC BY 4.0)opentruth to make
no truth on both images; Propagate atlas labels and check on 10 embryos where two people label 5 organs
Not large; included for whole-specimen micro-CT between subjects. Licence checked on 2026-10-10 at https://zenodo.org/api/records/4559800; derived test pairs shareable: yes; commercial use: yes. IMPC licence statement is the site footer; embryo image files served from the IMPC portalPropagate atlas labels and check on 10 embryos where two people label 5 organs; landmarks (limb tips, eye lenses, heart apex) are easy and repeatable
Longitudinal dark-field mouse CT (LDM-CT)in vivo attenuation and dark-field CT of mouse lungs over 12 weeksmice and time points unverified3D lung masks per scanboth imagesaccountCC BY 4.0openneeds an account
the download needs an account
Talbot-Lau lab system; healthy, injury and COPD models.airway branch points found by a program, checked by a person
Longitudinal multimodal MRI of rat ischemic strokeRat brain MRI over time, same animalUnverifiedUnverifieddirectnone stated (Zenodo record open, no licence field; paper says raw data made available for use)unknowntruth to make
42 rats with DWI, ADC, T2 and perfusion maps at occlusion, 30 min and 4 days after recanalisation; no lesion masks or landmarks; no licence stated
Licence checked on 2026-10-10 at https://zenodo.org/records/8139567; derived test pairs shareable: unclear; commercial use: unclear. Paper (JCBFM 2023, PMC10993873) is CC BY but says nothing on data terms beyond the Zenodo DOI.Check the content first; then two people place 10 landmarks (ventricles, corpus callosum, hippocampus points) away from the lesion on 10 time pairs.
Longitudinal T2 MRI of rat stroke treated with SAHAquantitative T2 maps of rat brain at day 1 and day 8 after transient MCAO, same animalanimal count unverifiednoneone imagedirectCC BY 4.0opentruth to make
no truth on both images; two people place 10 landmarks per rat
Lesion and oedema change between visits.two people place 10 landmarks per rat; lesion masks drawn on both days
Longitudinal young adult C57BL6J mouse MRI (Zenodo 6844489)11.7T mouse brain MRI, 101 labelled scans; inter-subject (longitudinal in title, single time point in text)101 scansLabels from earlier studies (manual or atlas: not stated)both imagesdirect: Zenodo record 6844489CC BY 4.0 (Zenodo)openleft out
Zenodo (CC BY 4.0, 0.36 GB): 101 T2 mouse-brain MRIs at 11.7 T, mice 30 to 44 scanned on two to four dates in May and June 2019, with brain-extraction masks only (label_adult.zip, 0.66 MB for all), the training data of the BEN skull-stripping network. A whole-brain mask measures nothing inside the brain, no public network labels mouse-brain structures for an automatic truth, and one contrast per session gives no pair aligned by construction for a known warp.
Check label method and time points.
LSPIG (liver CT of pigs with pneumoperitoneum, Harbin)pig abdominal CT, same animal without and with 13 mmHg pneumoperitoneum17 pigs, 34 directed pairsliver segmentations (made by the providing hospital)both imagesdirectnone stated for the data (repository code MIT; data on Google Drive)unknownleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): its licence is not shown to allow commercial use (none stated for the data (repository code MIT; data on Google Drive)). It was proposed as: Pig CT with and without pneumoperitoneum (17 pigs) with liver masks from the hospital: a setting the corpus lacks, but only 128^3 preprocessed crops and no data licence stated.
Only the preprocessed 128^3 crops are available, from the RCN repository links. Realistic intra-subject surgical deformation. Licence checked on 2026-10-10 at https://github.com/microsoft/Recursive-Cascaded-Networks; derived test pairs shareable: unclear; commercial use: unclear. README links the preprocessed SLIVER/LiTS/LSPIG crops with no terms; MIT covers the code in the repo, not the Drive data.two people mark 10 vessel bifurcations on both scans of each pig; liver masks stay as the overlap check
Ma 2008 in vivo C57BL/6J mouse atlasesIn vivo 9.4T mouse MRI 100 µm; inter-subject12 individual atlases = 132 directed pairsManual labels per mouse (individual atlases)both imagesrequestCC BY-NC-SA 4.0 (NITRC project c57bl_mr_atlas)non-commercialneeds a request
the data needs a request or a data use agreement
Host not confirmed (sister set of MRM NeAt). Licence checked on 2026-10-10 at https://www.nitrc.org/projects/c57bl_mr_atlas/; derived test pairs shareable: yes; commercial use: no. Same NITRC project hosts Ma 2005 (MRM NeAt, PMID 16165303) and Ma 2008 (PMID 18958199); ShareAlike applies to derived pairs.
Macaque MRI before and after deep brain stimulation, with histology (DANDI 001785)T1w and diffusion MRI of macaques before and after DBS, same animal; IF and EM histology3 monkeys with structural pre/post MRI (3 more with fMRI only)nonedirectCC BY 4.0opentruth to make
no truth on both images; electrode tracks as markers
Nature Neuroscience 2026 (DBS white matter remodelling)electrode tracks as markers; two people place 20 macaque brain landmarks on each pre/post pair
Marmoset brain registration benchmark (Zenodo)marmoset brain images and copies moved by known affine and translation transformscount unverifiedknown applied transformsboth imagesdirect: Zenodo, one zip readable by ranges, no account (originals also on the RIKEN CBS portal, doi 10.60178/cbs.20230630-001)CC BY 4.0 (Zenodo; derived from RIKEN CBS data under CC BY 4.0)openin the corpuspriority 3
The zip (1.5 GB, CC BY 4.0) holds 52 marmoset brains of the Brain/MINDS Marmoset Connectivity Resource and its template, each as the original serial two-photon tomography volume in its own space (0.1 mm, about 217 x 267 x 344) with an atlas label map, plus copies moved by a known affine or a translation. The moved copies leave nothing once the corpus aligns them, but the originals give real pairs between animals; their labels are the atlas mapped into each brain by the BMCR pipeline (ANTs registration to its STPT template), so automatic.
Synthetic affine only; derived from RIKEN Brain/MINDS connectivity data.
MIRACL CLARITY + MRI stroke miceCLARITY light-sheet and in vivo MRI of the same stroke mice to Allen atlas10 stroke mice (CLARITY, MRI, CT)~30 manual landmarks per volume by one author; TRE 130 um (CLARITY-ARA), 418 um (MRI-ARA)both imagesrequestno data licence; subset released as pipeline examples, rest available upon requestaskneeds a request
the data needs a request or a data use agreement
Only a subset released with the pipeline; rest on request. Fails access. Licence checked on 2026-10-10 at https://pmc.ncbi.nlm.nih.gov/articles/PMC6890789/; derived test pairs shareable: unclear; commercial use: unclear. Data availability statement of Goubran et al. Nat Commun 2019; GitHub repo AICONSlab/MIRACL licence not recognised by GitHub (NOASSERTION).Ask for the 10 mice and landmark files.
MIRCen macaque segmentation set (MACSET)7T T2 MR, 10 cynomolgus macaques; inter-subject10 = 90 directed pairsManual labels of 17 regions on 15 coronal sections per scan (sparse)both imagesdirect: https://www.nitrc.org/frs/download.php/10184/mircen_macset.tar.gz (no login)CeCILL v2 on the NITRC page, with 'freely available for academic work upon citing the Data in Brief article'askleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): its licence is not shown to allow commercial use (CeCILL v2 on the NITRC page, with 'freely available for academic work upon citing the Data in Brief article'). It was proposed as: Ten 7 T T2 MRIs of cynomolgus macaques in one 60 MB tarball that NITRC serves without a login. The manual labels cover 17 regions on 15 sections per scan in several orientations (NITRC page), sparse on both images, so warped labels cannot be scored; the corpus would have to make a truth (a macaque segmenter's labels on both brains, or a known warp), the sparse sections kept as a check. The page gives CeCILL v2 but also 'freely available for academic work upon citing', so the terms are unclear.
Sparse 2D labels in 3D volumes.
MMB-MRI multimodal mouse brain (Science Data Bank)11.7T mouse brain MRI in 6 contrasts (QSM, R2*, T1, T2, T2*, magnitude), between mice47 annotated mice of 75refined atlas-derived anatomical labels per brainboth imagesdirect: Science Data Bank file tree API and china.scidb.cn/download?fileId=..., per file (about 9.8 MB a volume), no accountCC BY 4.0 (Science Data Bank record)openin the corpuspriority 1
75 C57BL/6J mouse brains at 11.7 T (0.07 mm, 165 x 230 x 135), six co-registered contrasts each (QSM, R2*, T1, T2, T2*, magnitude) for 47 of them; the README says the 50-region labels of these 47 (Dorr atlas consolidated) were manually corrected, those of the other 28 atlas-mapped only. The record says processing included affine and deformable registration, yet all files share one grid and the mean region Dice between three of the 47 label maps as delivered is 0.60 to 0.68 (brain mask 0.91 to 0.93), so the brains are not warped onto each other and room remains, to confirm after the corpus's affine alignment. CC BY 4.0, files downloadable one by one without an account (china.scidb.cn); segmentation 550 can then be derived from the same images.
Labels started from an atlas, so they share its registration bias.an expert corrects 10 structures on 10 brains
Mouse tibia repeated micro-CT for DVC precision (Sheffield, Dall'Ara)Repeated micro-CT scans of mouse tibiae, in vivo and ex vivo, no load between scans; same boneRepeated-scan pairs (zero strain), counts unverifiedZero displacement by design (repeat scans, with and without repositioning)both imagesrequestORDA record CC BY 4.0, but the images are not on ORDA (too large); contact the authoraskneeds a request
the data needs a request or a data use agreement
Same group: bovine bone virtual-displacement and repeat-scan studies (2014, 2015, 2019). Licence checked on 2026-10-10 at https://orda.shef.ac.uk/articles/online_resource/MicroCT_Dataset_of_mouse_tibiae_for_DVC_precision_analyses_repeated_scans_ex_vivo_and_in_vivo_/5528104; derived test pairs shareable: unclear; commercial use: unclear. Record licence is open; confirm the files come under it when requesting them from e.dallara@sheffield.ac.uk.Zero-strain pairs are exact truth for non-rigid methods (rigid repositioning only); add a known warp to one scan for a non-zero test
Mouse tMCAo longitudinal MRI with lesion segmentationsmouse brain MRI after transient middle cerebral artery occlusion, several time points, same animallongitudinal series (count unverified)semi-automated and automated lesion segmentationsZenodo 4086450, Dryad's open mirror: direct, zips readable by ranges; Dryad itself serves files behind a browser captcha and its API needs a tokenCC0 1.0 (Zenodo and Dryad)openleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): open, but not among the six minor sets kept. It was proposed as: T2 maps of 121 scans of mice after tMCAo (Leiden set: 58 mice scanned at 4 h, 24 h, 48 h and 8 d, 28 of them at two to four time points; 128x128x16 voxels of 0.12x0.12x0.5 mm), with lesion masks by two observers using a semi-automated method and automated masks (data report, Zenodo listing). The only truth on both images is the lesion, which grows and shrinks between time points (to be shown, never ranked, as MSLesSeg's lesions), and the MHD headers carry no scanner frame (identity matrix, zero offset). A ranked truth needs landmarks placed by people or a known warp between two echoes of one T2 scan (synthetic).
also on Zenodo (record 4086450); lesions evolve, so masks are not a fixed referencetwo people place 15 landmarks away from the lesion on 10 animal pairs; atlas labels on both as a weaker check
Mouse-MammaryGenetically engineered mouse breast-cancer models: serial MRI32 mice, 149 studies (15 with 3-14 sessions)Noneone imagedirectCC BY 3.0opentruth to make
no truth on both images; Two people place 10 landmarks (spine, kidney poles, bones) on 10 serial pairs
Fails truth.Two people place 10 landmarks (spine, kidney poles, bones) on 10 serial pairs; tumours grow, so score outside them.
MRH mouse brain MRI + histologyMulti-contrast ex vivo MRI with co-registered histology (Nissl, NF, MBP)Unverified countCo-registration by authors (to Allen space)directCC0opentruth to make
ex vivo MRI of 20 mice delivered already warped into Allen space, histology (AF, NF, MBP) taken from the Allen atlas, not the same mice; only a known warp between same-session contrasts would give a truth
Already registeredMRI and histology are delivered co-registered to Allen space: no usable truth unless native images are released.
MRI and optoacoustic tomography registration (Klohs)MRI vs MSOT of fiducial phantom, ex vivo and in vivo mouse brainFew pairsFiducial markers in phantomdirect: Zenodo record 1303386CC BY 4.0 (Zenodo)openleft out
Zenodo (CC BY 4.0, 0.67 GB): one fiducial phantom and one ex vivo mouse brain as raw Bruker MRI folders with their MSOT data, one in vivo mouse brain as two TIFFs (MRI and MSOT HbO2), and the RegOA application. Three pairs in all, no annotation file, and only the phantom's fiducials could give a truth, for a rigid problem.
Phantom fiducials seen on MRI and MSOT give marker truth; for the mouse brains, two people mark 6 vessel points on both. Few pairs.
MRM NeAtMR microscopy of the mouse brain, inter-animal1020 structures by expertsboth imagesaccountCC BY-NC-SA 4.0 (NITRC project c57bl_mr_atlas)non-commercialneeds an account
the download needs an account
Licence checked on 2026-10-10 at https://www.nitrc.org/projects/c57bl_mr_atlas/; derived test pairs shareable: yes; commercial use: no. Ma et al. 2005 ex vivo atlas, shared with the 2008 in vivo atlas on NITRC, no account needed for the page; ShareAlike.
MSOT and MRI of murine tumours (Cambridge, 2019)multispectral optoacoustic tomography to 9.4T MRI, same mousemice count unverifiedmanual registration landmarks plus tumour and body contours in bothboth imagesdirectCC BY 4.0openwaits for a tool
single-slice 2D optoacoustic images matched to an MR slice.
Different from the Klohs brain set already listed.
Multiscale heart micro-CT + SBF-SEMWhole-heart micro-CT and SBF-SEM blocks of same sample4 heartsNoneone imagedirectCC0opentruth to make
no truth on both images; Two people place 10 landmarks (vessels, capillary branches) on the micro-CT region and the SBF-SEM block of each of the 4 hearts
Correlative; truth failsTwo people place 10 landmarks (vessels, capillary branches) on the micro-CT region and the SBF-SEM block of each of the 4 hearts; a demo case, not a test set.
PDMR-Texture-Analysis (+ PDMR-* model subsets)Patient-derived xenograft mice: serial non-contrast T2w MRI, same animal over weeks175 mice, 689 studies (2-9 per mouse)None seen (SR reports, no masks confirmed)one imagedirectCC BY 4.0opentruth to make
no truth on both images; Two people place 10 landmarks (spine, kidneys, bones) on 10 serial pairs
Fails truth.Two people place 10 landmarks (spine, kidneys, bones) on 10 serial pairs; tumours grow, so score outside them.
Photoacoustic mesoscopy co-registration dataset (Cambridge)3D photoacoustic mesoscopy of breast cancer patient-derived xenografts imaged twice (same animal, two sessions)count unverifiedvessel segmentations on both images (automatic, method in paper)both imagesdirect: Apollo, University of Cambridge repository, no accountCC BY 4.0 (item page 'Rights and licensing'; README.docx)openleft out
The Apollo record (doi 10.17863/CAM.111991, CC BY 4.0) holds a single pair: one breast cancer patient-derived xenograft imaged twice, as preprocessed fixed and moving volumes with their vessel segmentations (preprocessedImages.zip), the two raw reconstructions (1.1 GB .mat each) and the five methods' warped results. One pair cannot make ten test pairs, and its vessel segmentations come from the authors' segmentation pipeline (automatic, per the screening row), not from people.
Not aligned by construction: a real repeat-imaging registration set, found while searching.two people place 20 vessel bifurcations on each pair; segmentations give a vessel-overlap check
Porcine liver multi-phase CE-CT (RFA study, ISBI 2014)three contrast phases (hepatic artery, portal vein, hepatic vein) of porcine livers at respiratory arrest, same animal; some with induced lesions57 livers x 3 phasesnone stated (vessel trees can be reconstructed)requestnone stated (data were linked from gosmart-project.eu/datasets.php, now defunct)unknownneeds a request
the data needs a request or a data use agreement
Download link not found; access unverified (Kainz et al.). Licence checked on 2026-10-10 at http://web.archive.org/web/20141115130514/http://gosmart-project.eu/datasets.php; derived test pairs shareable: unclear; commercial use: unclear. Paper (Kainz et al. ISBI 2014, Spiral 10044/1/27850) and the archived GoSmart page list Pig01-Pig98.rar with no terms; host gone.vmtk bifurcations of vessels seen in two phases, matched by a program and verified by a person on 10 livers; access unverified.
Preclinical CBCT with BioXmark injectable fiducial (mouse)small-animal CBCT of mice with an injected liquid fiducial marker6 files; repeated scans of the same mouse unverifiednone; fiducial marker visible in the imagesdirectCC BY 4.0opentruth to make
no truth on both images; fiducial centroid per scan as a landmark if repeat scans exist
accompanies a feasibility study of BioXmark for preclinical radiotherapy targetingfiducial centroid per scan as a landmark if repeat scans exist; otherwise only a known warp
PRIME-DEMacaque MRI from 25+ sites, 217+ animals; inter-subject, some longitudinalManyNonedirectPer collection, mostly CC BY-NC-SAnon-commercialtruth to make
no truth on both images; Two raters place 20 landmarks on 10 same-animal or inter-subject macaque pairs
Fails truth. status change: Round 4 count: listed in 2 of the 38 round-4 surveys and benchmark papers (LUMIR 2025; LUMirage 2025).Two raters place 20 landmarks on 10 same-animal or inter-subject macaque pairs; mostly non-commercial licences.
Rat multi-atlas (Lancelot 2014)7T T2 MR, 12 Sprague-Dawley rats; inter-subject12 = 132 directed pairsManual 27-region labels per ratboth imagesrequestavailable with grant of licence from CERMEP, University of Lyonaskneeds a request
the data needs a request or a data use agreement
Host not confirmed. Licence checked on 2026-10-10 at https://pmc.ncbi.nlm.nih.gov/articles/PMC4201469/; derived test pairs shareable: unclear; commercial use: unclear. Paper is CC BY but the atlas dataset is licensed separately by CERMEP.
SIGMA rat longitudinal EPI setRat EPI/diffusion MRI, 35 rats x 5 time points~140 longitudinal pairsNoneaccountFree for non-commercial use only (NITRC licence field)non-commercialneeds an account
the download needs an account
Fails truth. Licence checked on 2026-10-10 at https://www.nitrc.org/projects/sigma_epib0/; derived test pairs shareable: unclear; commercial use: no. NITRC licence says nothing on redistribution.Two raters place 10 landmarks on 10 rat time pairs (EPI distortion makes them coarse); cycle consistency across the 5 time points.
ULMShare (transcranial ultrasound localization microscopy in mice)raw ultrafast ultrasound of mouse brain with microbubble localisations and tracks99 acquisitions from 61 micemicrobubble trajectories from the processing pipeline (not independent)one imagedirectunverifiedunknowntruth to make
no truth on both images; simulate microbubbles in a known vascular phantom for exact tracks
Motion correction and frame registration within acquisitions; tracks are algorithm output.simulate microbubbles in a known vascular phantom for exact tracks
UNC-Wisconsin rhesus neurodevelopment MRILongitudinal macaque T1, 2 weeks to 3 yearsCohort, multiple time pointsNoneone imagedirectCC BY 4.0opentruth to make
no truth on both images; Two raters place 15 landmarks on 10 same-animal pairs across development
Two raters place 15 landmarks on 10 same-animal pairs across development; growth limits label overlap to a weaker check.
UW-CIRP-Mouse-PET-CT-NSCLC (IDC)Genetically engineered mouse model of lung squamous carcinoma: 18F-FDG PET/CT before and after combination therapy, same animal11 mice x 2 sessions (+3 baseline-only)RTSTRUCT segmentation per session (structures not checked)both imagesZenodo manifests; files in IDC's public buckets, no loginCC BY 4.0 (Zenodo)openleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): open, but not among the six minor sets kept. It was proposed as: 11 GEMM mice with lung squamous carcinoma imaged with FDG PET/CT on a Siemens Inveon before and after combination therapy (plus 3 baseline only); CT 1024x1024x1345 at 0.099 mm, 1.41 GB of DICOM a CT, 35.6 GB in all (Zenodo record, IDC bucket listing). Each session has a manual RTSTRUCT (PMOD) with the lesions (three on the one read), a liver ROI and a blood-pool ROI: tumours that change under therapy and SUV references, so no ranked truth across sessions, and 11 animals leave no development set; a known warp between the PET and CT of one session (hybrid scanner) is what the corpus can make.
IDC-only (Co-CIRP), published 2026-05. Structures segmented: unverified.
Waxholm Space C57BL/6 mouse (Johnson 2010)Ex vivo mouse MR microscopy 21.5 µm (T1/T2/T2*), 14 specimens, Nissl on a subset registered to MRI14 specimens; MR-Nissl for a few37-structure labels (probabilistic atlas); per-specimen manual labels unverifiedaccountCC BY-NC-SA 4.0 (INCF WHS atlas on NITRC)non-commercialneeds an account
the download needs an account
Check if per-specimen labels exist. Licence checked on 2026-10-10 at https://www.nitrc.org/projects/incfwhsmouse/; derived test pairs shareable: yes; commercial use: no. Canonical WHS T1/T2*/T2 volumes and labels; per-specimen CIVM VoxPort data may carry other terms (registration).T1, T2 and T2* of one fixed specimen are aligned by construction: a known warp gives exact multi-contrast truth; inter-specimen pairs get 15 landmarks by two people on 10 pairs.
Waxholm Space Sprague Dawley rat brain atlas (v4)Ex vivo MRI (39 um) and DTI (78 um) of one rat brain with 222 delineated structures; templateNone (single template; target for rat section and volume registration)222-structure manual delineations on the templateone imagedirectCC BY 4.0opentruth to make
no truth on both images; Truth must come from subject data: two people place 20 landmarks on rat sections and the template
Template, not pairs; pairs need subject images (e.g. EBRAINS rat sections)Truth must come from subject data: two people place 20 landmarks on rat sections and the template
Whole zebrafish X-ray histotomography (Ding 2019)Synchrotron micro-CT of whole fixed zebrafish (larval to juvenile) at sub-micron to 1.43 µm; between fish (5 dpf larvae), one specimen per stageSeveral 5 dpf fish registered together; development seriesBrain-region segmentations and nuclear detections (authors); registration pipeline includedone imagedirectCC0 (Dryad API)opentruth to make
no truth on both images; Nuclei centroids matched between fish are not 1:1
Nuclei centroids matched between fish are not 1:1; use region labels from two people on 5 fish, or a known warp at full resolution

Phantom · 80

DatasetProblemPairsChecked againstAccessLicenceStatusNotesTo make the truth
3D-printed deformable female pelvis phantom for EBRT to brachytherapy DIR (Kadoya 2017)pelvis phantom CT without applicator (external beam) and with applicator (brachytherapy), deformable uterus and bladdera few phantom statesspatial DIR accuracy measured on phantom features (details unverified)hiddennot releasedaskleft out
its images or its truth are not released
Not public.request from Tohoku
4D CBCT dynamic thorax phantom (Madesta)4D CBCT of CIRS 008A with programmed insert motion6 scansKnown motion pattern (+/-10 mm SI)both imagesdirectCC BY-NC-SA 4.0non-commercialwaits for a toolpriority 3
The zip holds raw projections only (raw_data/*.zip per session): needs a 4D CBCT reconstruction before any registration.
4DCT XCAT phantom with irregular breathing (UCL, MICCAI 2024)simulated unsorted 4D CT of the XCAT phantom with variable breathing1 simulation, many phasesXCAT motion is known by construction; whether displacement fields are shipped is unverifiedboth imagesdirectCC BY-NC-ND 4.0non-commercialtruth to make
one XCAT simulation, the motion fields are not stated to be in the files; ND licence.
single example released with Paul Segars' permission; ND licence forbids redistributing modified dataexport XCAT organ masks and the phantom's motion vectors per phase for exact point truth
AAPM TG-132 datasetsdigital and physical phantoms for registration QAa fewknown transformsboth imagesdirectAAPM MPEC disclaimer: research purposes only, not for patient care; no licenceaskneeds an account
the download now needs an AAPM login; its deformations come from a commercial tool.
status change: the dataset page asks for an AAPM login (listed as direct, no URL). The 4D CT is a DIR-Lab case; deformations come from the commercial ImSimQA tool, so they favour similar smooth models. Licence checked on 2026-10-10 at http://web.archive.org/web/20211208223858/https://www.aapm.org/pubs/reports/report132.asp; derived test pairs shareable: unclear; commercial use: unclear. Download was behind an I-agree disclaimer; the live page now asks for an AAPM login.
ADAM-pelvis anthropomorphic deformable CT/MR phantom (DKFZ)male pelvis phantom with deformable silicone organs, CT and MRI, varying bladder and rectum fillingphantom states; image count not releasedimaging markers in the organ shells; dosimeter pocketsboth imageshiddennot releasedaskleft out
its images or its truth are not released
Physical phantom; images not shared as far as found.request scans from DKFZ; markers give multimodal point truth
ASTRA Toolboxsimulator (software): pairs are generated, not downloaded; 2D/3D tomography forward projection and reconstruction on GPUCT vs simulated limited-view/CBCT reconstructionsknown geometry and object by constructionboth imagesdirectGPL-3.0openleft out
simulator software, no images to download (synthetic truth like the known-warp track's).
copyleftsame as TIGRE: warp a CT by a known field, project and reconstruct
BigBrain-MR (100 um digital MR phantom)BigBrain histology mapped to MR properties: T1w, T1 map, T2*w, R2*, QSM, B0, coil maps, 20-ROI atlasmulti-contrast maps of one brain on one grid (100-800 um)all maps aligned by construction; atlas labelsboth imagesdirectCC BY-NC-SA 4.0non-commercialtruth to make
one subject with maps aligned by construction, so only a known warp; 34 GB; NC-SA.
one subject; identity alignmentsimulate two contrasts, warp one with a known field: exact multimodal truth at very high resolution (large-volume test)
BrainWeb 20 normal anatomical models20 different normal brains, 12-class fuzzy and discrete tissue models; simulated T1 (SFLASH) per subject190 between-subject pairs; within-subject multi-contrast via the brainweb Python package (PET/MR, T1/T2)exact tissue labels on every subject (12 classes)both imagesdirectnone stated on the BrainWeb pagesunknownin the corpuspriority 3
In the corpus: the 20 phantoms' simulated T1 paired into 10 test pairs, no development, the tissue classes as truth (synthetic), ranked apart.
only T1 offered on the site for these models; pip package casperdcl/brainweb builds more contrasts Licence checked on 2026-10-10 at https://brainweb.bic.mni.mcgill.ca/brainweb/anatomic_normal_20.html; derived test pairs shareable: unclear; commercial use: unclear. Main page, model page and FAQ carry no terms; the brainweb pip package (MPL-2.0) covers code only.between-subject registration scored with Dice on the 12 labels; or warp one subject's simulated T1/T2 by a known field for exact truth
BrainWeb Simulated Brain Database (normal and MS)simulated T1, T2 and PD brain MR from one anatomical model (normal or MS), many slice thicknesses, noise and non-uniformity levelsT1/T2/PD of the same anatomy at any noise/INU setting; custom simulations on the serverexact alignment across contrasts and tissue/lesion labels by constructionboth imagesdirectnone stated on the BrainWeb pagesunknowntruth to make
one simulated anatomy aligned across contrasts, so only a known warp; same data as 'BrainWeb simulated brain MRI'; no licence stated.
single anatomy (Colin27-based); identity alignment, so a deformation must be added Licence checked on 2026-10-10 at https://brainweb.bic.mni.mcgill.ca/brainweb/; derived test pairs shareable: unclear; commercial use: unclear. No licence or terms text on the main page or FAQ.apply a known smooth warp to one contrast (e.g. T1 vs warped T2) for exact multimodal truth; vary noise/INU to test robustness
CBCT registration accuracy phantom study (Mendeley)CBCT of a phantom under displacements, artefacts, thickness changes and protocolscount unverifiedapplied displacements (unverified)directCC BY 4.0openleft out
no images: the record holds one spreadsheet of registration results
Description is only the title.known couch shifts give rigid truth
CBICA AtrophySim (Karacali & Davatzikos 2006)topology-preserving warp that produces a prescribed volume loss in a labelled regionany baseline/follow-up pair from one input MRIdense deformation with prescribed Jacobian, exact correspondencesboth imagesdirectSBIA License: free for non-commercial research only, no redistributionnon-commercialleft out
software that warps a scan (synthetic atrophy); non-commercial, no redistribution.
warps are smooth and model-driven, may favour some regulariserssimulate atrophy on real scans, add independent noise/rescan to avoid identical texture
Colin27 average brain 2008 (MNI)one subject averaged: T1 (27 scans), T2 (11), PD (12) on one 0.5 mm grid, plus discrete and fuzzy tissue modelsT1/T2/PD of one anatomy, alignedcontrasts aligned by the averaging registration; 12 tissue classesboth imagesdirectcopyright Louis Collins/MNI: use, copy, modify and distribute for any purpose without fee, keep the noticeopentruth to make
one subject's contrasts aligned by averaging, so only a known warp.
permissive licence; identity alignment onlywarp T2 or PD by a known field and register to T1: exact multimodal truth at 0.5 mm
CONRAD (cone-beam framework, FAU)simulator (software): pairs are generated, not downloaded; generates 4D projection and volume data with their vector motion fields4D CBCT phases, projections with known motionvector motion fields output by the simulatorboth imagesdirectGPL-3.0 (contains NVIDIA code)openleft out
simulator software, no images to download.
Java; XCAT-based scenes need XCAT files (unverified)simulate a moving phantom, export the motion fields as truth for 4D-CBCT/CT registration
CT4Harmonization-MulticentricAnthropomorphic phantom, 13 scanners, 18-30 repeats1,646 seriesSame rigid phantom (identity anatomy)direct: TCIA's NBIA API or Data Retriever manifest, no accountCC BY 4.0 (TCIA page)openleft out
One anthropomorphic phantom (1 subject) scanned on 13 scanners at five doses, 1,378 series in 252.38 GB: two scans differ by a rigid shift and voxel spacing, which the corpus's rigid alignment already removes, leaving no deformation to register. Every pair is the same object, so test and development sets cannot hold different subjects.
Spatial shifts preserved Licence checked on 2026-10-10 at https://www.cancerimagingarchive.net/collection/ct4harmonization-multicentric/; derived test pairs shareable: yes; commercial use: yes. TCIA data access table, version 1 (2025-06-09).The phantom is rigid: two people mark 10 distinct phantom features per scan, and a rigid fit gives exact truth across scanners and repeats.
DECT and photon-counting CT phantom scans (Lustermans, Maastricht)tissue-equivalent and anthropomorphic phantoms on Siemens NAEOTOM Alpha (PCCT) and SOMATOM Confidence (DECT), several dose levelsfew phantoms x 2 scanners x 3 dose levels (three Zenodo records: 18377176, 17119205, 15119102)none; known phantom geometrydirectCC BY 4.0opentruth to make
no truth on both images; cross-scanner: fit a rigid transform from phantom inserts (insert centres as landmarks)
Within one scan the energy images (VMIs, bins) are aligned by construction. Across scanners the phantom is rigid, so the true transform is rigid.cross-scanner: fit a rigid transform from phantom inserts (insert centres as landmarks); within scan: warp one energy image and register another
DeepDRR (realistic DRR with scatter and noise)simulator (software): pairs are generated, not downloaded; physics-based X-ray simulation from CT (material decomposition, scatter, noise)unlimited CT-to-X-ray pairs, tools can be insertedpose and tool positions known by constructionboth imagesdirectGPL-3.0openwaits for a tool
X-ray simulator for 2D/3D pairs (waits for the 2D/3D tool).
GPL copyleft if redistributed inside a servicesimulate fluoroscopy at sampled C-arm poses from public CTs for 2D/3D truth
Depth sensors for markerless neurosurgical registrationDepth-camera point clouds of a head phantom vs reference modelSeveral sensors x sessionsCalibration and reference registrationsdirectCC BY 4.0openwaits for a tool
depth-camera point clouds of a head phantom against an MRI-derived point cloud, with reference transforms and TRE files: waits for mesh and point-cloud scoring
Point clouds, not image pairsPoint clouds, not images: the calibration with phantom fiducials is the truth, out of scope for image pairs.
DiffDRR (differentiable DRR, PyTorch)simulator (software): pairs are generated, not downloaded; DRRs from any CT at any 6-DoF pose; base of DiffPose and xvr 2D/3D registrationunlimited CT-to-X-ray pairs at chosen posespose known exactly by constructionboth imagesdirectMITopenwaits for a tool
X-ray simulator for 2D/3D pairs (waits for the 2D/3D tool).
simple attenuation model, no scatterrender DRRs of public CTs at sampled poses as 2D/3D truth; test on real fluoro (DeepFluoro) to measure the realism gap
Digimouse (USC 3D mouse atlas)one mouse: CT (0.1 mm), PET (FDG), cryosection colour images, co-registered, with 21-structure atlas and meshesCT-PET-cryosection of one mouseco-registration by the authors and atlas labelsboth imagesdirect: zips on neuroimage.usc.edu (co-registered set about 77 MB, raw CT 900 MB, cryosections 600 MB)CC BY-SA 4.0 (Digimouse_license page)openleft out
One animal: a 28 g nude male mouse, its CT, cryosection and PET co-registered by the authors, with 21 atlas structures (Digimouse page). Direct download under CC BY-SA 4.0 (licence page), but one mouse gives one case at most, and test and development cases need different subjects.
alignment is a registration result, not exactwarp one modality by a known field for exact truth; or two people place 20 landmarks on CT and cryosection
DIREPvirtual head and neck phantoms with known deformation10known deformation (synthetic)both imagesdirectvirtual phantom CTs free to download with no licence; ground-truth DVFs and contours sold by Oncology Systems Limitedaskleft out
the truth (DVFs and contours) is sold, not released.
check the source status change: host page found; the truth is hidden (upload-and-score, DVFs for sale), not direct as listed. Same data as the listed 'Virtual head and neck phantom benchmark (JACMP 2021)'. Whether the scoring service still answers is unverified. Licence checked on 2026-10-10 at https://sites.google.com/site/dirphantoms/home; derived test pairs shareable: unclear; commercial use: unclear. Scoring service closed to new submissions since 2018-03-01; truth must be bought from OSL, whose terms are not public.ask OSL or the authors to release DVFs; or warp the start images with a second known smooth model unlike the ones tested
Duke 4D beating-heart library from PROMISE 4D cardiac CT4D heart mesh models (chambers, myocardium, coronaries) with frame-to-frame motion from LDDMM on 32 ECG-gated 4D CT (10 frames + 1 high-resolution frame)32 hearts x continuous cardiac cycleknown motion of the model by construction (derived from AI segmentation + LDDMM)both imagesrequestto be released like XCAT phantoms (XCAT licence, unverified)askneeds a request
the data needs a request or a data use agreement
original PROMISE 4D CT not released; models plug into XCAT and DukeSim to simulate CT with exact motionsimulate 4D CT with DukeSim from the models: exact displacement truth
DukeSim (Duke CVIT CT simulator)simulator (software): pairs are generated, not downloaded; GPU CT simulator (ray tracing + MC scatter, PCD, dual energy) for voxel phantomsCT of XCAT phantoms at any state/protocolphantom motion and labels knownboth imagesrequestrequest through the CVIT REDCap form; terms not shown publiclyaskneeds a request
the data needs a request or a data use agreement
pairs naturally with XCAT 3.0 Licence checked on 2026-10-10 at https://cvit.duke.edu/resource/dukesim-v1-2/; derived test pairs shareable: unclear; commercial use: unclear. Form is captcha-gated; no licence text on the resource page.simulate two breathing states of an XCAT 3.0 phantom, XCAT vectors are the truth
Duplex simulated echocardiography (CREATIS/IMAG)simulated echocardiographic duplex (B-mode + Doppler) sequences20 sequencesreference motion by constructionboth imagesdirectnone stated (page calls it an open-access database)unknownwaits for a tool
2D B-mode and colour Doppler frames; the truth is simulated blood flow.
details unverified (only the database listing seen) Licence checked on 2026-10-10 at https://humanheart-project.creatis.insa-lyon.fr/duplex.html; derived test pairs shareable: unclear; commercial use: unclear. No licence file in the DUPLEX Girder collection (code zip, data folder).
Dynamic anthropomorphic thorax phantom: 4D CT and Halcyon HyperSight CBCT with patient breathing traces (Maastro)dynamic thorax phantom scanned on 4D CT (Siemens) and CBCT (Varian Halcyon HyperSight) under patient-specific breathing tracesseveral traces, each with 4D CT and CBCT (count unverified)the programmed breathing traces (RPM curves included); no landmarks stateddirect: Zenodo record 17256015 (processed.zip 0.86 GB, readable by ranges; DICOM.zip; breathing_curves.zip; README.pdf)CC BY 4.0 (Zenodo)opento build apartpriority 3
Zenodo (CC BY 4.0, 1.6 GB): one 3D-printed breathing thorax phantom driven by five patient traces, 38 fast (6 s) or slow (60 s) HyperSight CBCTs rigidly registered by the provider to the 3D 4DCT reconstruction of their trace (six 4DCTs, 512x512x127, NIfTI), with one manual delineation per image: a single label of about 5,000 voxels, the tumour insert. One object and one small structure: ranked apart, and thoracic CBCT/CT registration is covered on patients (L2R_ThoraxCBCT).
Maastro Clinic (Daenen, Lustermans et al.). Phantom model and insert motion unverified.use the known insert motion from the traces as truth; mark the insert or embedded features in every phase
Dynamic respiratory motion phantom MRIradial MRI of a physical phantom moving 1D sinusoidally (10-40 mm peak-to-peak, 4 s period)4 scans (4 amplitudes)programmed motion amplitude and periodboth imagesdirect: Zenodo record 19744580CC BY 4.0 (Zenodo)openleft out
Zenodo (CC BY 4.0, 4.4 GB): four radial MRI acquisitions of a phantom moving in 1D sinusoids of 10 to 40 mm, delivered as raw k-space only (kspace.mat, 0.9 GB each) with reconstruction code and a network (ForErik_code.zip). Rigid 1D translation of a phantom, no images, no annotations: nothing to score.
rigid 1D translation only; data may be raw radial k-space (unverified)
Field IIsimulator (software): pairs are generated, not downloaded; linear acoustics ultrasound simulation from scatterer maps (Matlab/Python)US images of moving scatterers (any motion)scatterer displacement known by constructionboth imagesdirectcitationware; free to distribute unmodified; no use in commercial products without prior written consent of J.A. Jensennon-commercialleft out
ultrasound simulator software; commercial use needs written consent.
commercial use needs written consentmove scatterers with a known field (e.g. from an E/M heart or FE model) and simulate frames
FlowNet-PET training set (XCAT respiratory PET)300 XCAT phantoms with random anatomy and diaphragm motion, 10 PET activity frames over one breath300 x 10 frames (270 train, 30 val)same phantom over the breathing cycle by construction; motion vectors not included (unverified)direct: Zenodo record 6510358, one 3.3 GB zip holding one 48.7 GB HDF5CC BY 4.0 (Zenodo); code GPL-3openleft out
Zenodo (CC BY 4.0): 300 digital XCAT phantoms, ten simulated PET activity frames over one breath at 2x4x4 mm with count noise, in a single 48.7 GB HDF5 deflated inside one 3.3 GB zip, so nothing can be read by ranges and the record does not say whether motion vectors are included (the paper compares flows to ground truths). The images are simulations of a licensed phantom whose vectors would need the XCAT licence (on request) to regenerate; breathing is covered on real 4DCT with people's landmarks.
low resolution (2x4x4 mm) activity maps with count noiseregenerate the same XCAT parameters with the vectors program for dense truth (needs XCAT licence); otherwise use lesion position per frame
FOCUS (MSU ultrasound simulator)simulator (software): pairs are generated, not downloaded; fast nearfield pressure and transient field simulation (Matlab)US fields/B-mode of moving scatterersscatterer motion knownboth imagesdirectfree download with an as-is legal notice; no licenceaskleft out
ultrasound simulator software, no licence.
binaries only (unverified) Licence checked on 2026-10-10 at https://www.egr.msu.edu/~fultras-web/download.php; derived test pairs shareable: unclear; commercial use: unclear. A simulator, not data: pairs we simulate are our own output, but the tool itself has no stated licence.as Field II
FPUS23 fetus phantom ultrasound2D ultrasound of a 23-week fetus phantom (US-7 SPACE FAN-ST), probe moved around standard planes, four phantom orientations15,728 frames from sweeps; consecutive frames and repeated planesplane labels, orientation labels and anatomy bounding boxes; no probe pose recordedone imagedirectMIT (repository LICENSE; images on a Google Drive link from the README)opentruth to make
no truth on both images; phantom CT plus tracked re-acquisition would give exact poses
Static phantom: anatomy does not deform, so any two frames of the same plane differ only by probe motion; boxes are weak truth Licence checked on 2026-10-10 at https://github.com/bharathprabakaran/FPUS23; derived test pairs shareable: yes; commercial use: yes. Paper says the dataset is made public at this repository; the MIT text speaks of software, so the Drive images are covered only by implication.phantom CT plus tracked re-acquisition would give exact poses; meanwhile box centres as weak correspondences
GATE / OpenGATE 10 (Geant4 imaging and RT)simulator (software): pairs are generated, not downloaded; Monte Carlo PET, SPECT, CT, RT; moving and 4D voxel phantomsPET/SPECT/CT of the same moving phantomphantom motion and co-registration known by constructionboth imagesdirectLGPL-3.0openleft out
simulator software, no images to download (Monte Carlo PET/SPECT/CT).
PET/CT and SPECT/CT multimodal truthsimulate PET and CT of one phantom (XCAT, MOBY or patient-derived) at different breathing phases
gVirtualXray (gVXR)simulator (software): pairs are generated, not downloaded; GPU X-ray projection from polygon meshes (Beer-Lambert, polychromatic)radiographs of mesh anatomy at any posepose known by constructionboth imagesdirectBSD-3-Clause (per PyPI and Bangor pages; LICENSE file not read)openwaits for a tool
X-ray simulator for 2D/3D pose (waits for the 2D/3D tool).
meshes, no voxel texturerender meshes (e.g. ICRP 145) at known poses for 2D/3D pose truth
Henry Ford FEM patient phantoms (Stanley 2013)patient CT (4 lung, 1 prostate) deformed by finite-element models; simulated CBCT noise; physical lung phantom11 computational models plus a physical lung phantomFEM displacement fields; homologous points on the physical phantomboth imageshiddennot releasedaskleft out
its images or its truth are not released
Not public.request from Henry Ford; or FEM warps of public 4D-Lung/Pelvic CT
ICRP reference computational phantoms (Publications 110, 143, 145, 156)adult (110) and paediatric (143) voxel phantoms from CT; adult (145) and paediatric (156) mesh versionsbetween-subject pairs (male/female, newborn to adult); voxel vs mesh of the same personorgan labels by construction on every phantomboth imagesdirectICRP copyright; phantom files ship as supplementary material of the publications; no licence foundaskleft out
organ label phantoms, no images; no licence found.
dosimetry phantoms; no images, only organ/material labels Licence checked on 2026-10-10 at https://www.icrp.org/publication.asp?id=ICRP+Publication+110; derived test pairs shareable: unclear; commercial use: unclear. Supplement page on Sage is behind a bot wall; ICRP free downloads page says ICRP keeps copyright and materials must not be repackaged for sale.convert labels to CT numbers, deform with a known field (mesh phantoms support posture changes) for exact truth; or use between-subject with Dice
Image-to-Physical Liver Registration Sparse Data Challengepreoperative liver mask and mesh to sparse intraoperative surface points of a deformed phantom112 registration scenarios159 subsurface targets; blinded except 4 cases; scored via dashboardboth imagesdirectCC BY 4.0 (OSF project licence, 2023)openleft out
mesh to sparse surface points, and the targets are hidden.
Mesh-to-points, not image-to-image; truth hidden. Licence checked on 2026-10-10 at https://osf.io/u3dxy/; derived test pairs shareable: yes; commercial use: yes. Licence read from the OSF API for node u3dxy; cite Brewer et al. 2019 and Collins et al. 2017.
ImSimQA patient-based digital phantoms (Oncology Systems Limited)real patient CT (head and neck, thorax, prostate/pelvis) warped by known synthetic DVFs, same patientVaradhan 2013: 3 cases; Loi 2018: 3 sites x 2 DVFs; TG-132: about 10known DVF and warped reference contoursboth imageshiddencommercial softwareaskleft out
its images or its truth are not released
A generator, not a released set: the study images were not shared. The most used route to known truth in commercial DIR validation.apply known smooth warps to public CT (Pelvic-Reference-Data, HNSCC, 4D-Lung) with a model unlike the tested algorithms
In-house deformable anthropomorphic pelvis phantom, four DIR packages (Z Med Phys 2025)pelvis phantom CT, one static and four deformation scenarios, CT numbers overridden from a patient scan4phantom deformation (details unverified)hiddennot releasedaskleft out
its images or its truth are not released
Not public; details from the abstract only.request
JEMRISsimulator (software): pairs are generated, not downloaded; general Bloch MRI simulator with motion trajectories and coil modelsMR images of a phantom at known poses/motionmotion and alignment knownboth imagesdirectGPL-2.0openleft out
simulator software, no images to download (MRI).
slow for large 3Dsimulate the same phantom under a known rigid/nonrigid motion
k-Wavesimulator (software): pairs are generated, not downloaded; full-wave acoustic and photoacoustic simulation (Matlab, C++/CUDA)US/photoacoustic images of a medium in two statesmedium deformation knownboth imagesdirectLGPL-3.0openleft out
simulator software, no images to download (acoustics).
slow for B-mode at scalesimulate US of a medium before/after a known warp (photoacoustic-US multimodal too)
KomaMRI.jlsimulator (software): pairs are generated, not downloaded; GPU Bloch simulation, Pulseq sequences, arbitrary phantom motion (v0.9+)multi-contrast MR of the same phantom; moving phantomsphantom motion and alignment known by constructionboth imagesdirectMITopenleft out
simulator software, no images to download (MRI).
fast and permissive; phantoms from BrainWeb/XCAT etc.simulate T1w and T2w of a labelled phantom with a known motion: exact multimodal and motion truth
LUMIC (lung subtraction CT digital phantom)Simulated unenhanced vs contrast-enhanced chest CT (XCAT), same virtual patient, different breathing/cardiac phaseTest: 1 enhanced vs 3 diaphragm shifts (3/8/20 mm) + noise and titanium-pin variants; train: 2 pairsKnown dense displacement field from the XCAT phantom (organisers, Radboud)both imagesaccountnone stated on public pages; download needs a grand-challenge loginunknownneeds an account
the download needs an account
Rare: voxel-wise ground-truth DVF for multi-contrast lung CT status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Li 2023 dataset collection (ACM CSUR)). Licence checked on 2026-10-10 at https://lumic.grand-challenge.org/Dataset/; derived test pairs shareable: unclear; commercial use: unclear. Images were generated with the Duke XCAT phantom, whose licence may also bind them.
MC-GPU (VICTRE Monte Carlo X-ray)simulator (software): pairs are generated, not downloaded; GPU Monte Carlo X-ray transport for mammography, DBT, CBCT from voxel phantomsprojections of any voxel phantom stateknown geometry by constructionboth imagesdirectUS public domain (FDA disclaimer)openleft out
simulator software, no images to download (X-ray projections).
slow vs ray-tracing but realistic scatterimage compressed vs uncompressed VICTRE phantoms or warped XCAT for truth
MICCAI 2016 PET segmentation challenge (PETSEG)PET phantom and clinical PETunverifiedphantom geometryrequestno data page or licence foundunknownneeds a request
the data needs a request or a data use agreement
Listed in Li 2023 as phantom registration; fails pairs for registration. Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Li 2023 dataset collection (ACM CSUR)). Licence checked on 2026-10-10: no primary page states terms; derived test pairs shareable: unclear; commercial use: unclear. Challenge portal (FLI-IAM) no longer lists PETSEG; challenge paper Hatt et al. MedIA 2018 full text not retrievable from Europe PMC.phantom geometry gives exact truth if two scans exist
Michigan deformable lung phantom with 48 markers (Kashani 2008)deformable thorax phantom CT at inhale and exhale states1 (two or more deformation states)48 small plastic markers located manually, then erased from the images before sharingboth imageshiddennot releasedaskleft out
its images or its truth are not released
Listed in TG-132 talks as the 'Mich lung phantom'. Not public.request from Michigan; the design (markers then digital removal) is a model for blinded physical truth
MIDA head and neck model (IT'IS / FDA)multimodal-MRI-based head and neck model, 115 structures, 0.5 mm, voxel and STLlabel model only (one subject); images must be simulatedlabels by construction; raw MR/DTI not distributedboth imagesrequestIT'IS MIDA v1.0 user licence: free, non-exclusive, non-transferable; model data and derived works must not be distributedaskneeds a request
the data needs a request or a data use agreement
fails for direct use: no images; one subject Licence checked on 2026-10-10 at https://itis.swiss/assets/Downloads/VirtualPopulation/License_Agreements/LicenseAgreementMIDA_2024.pdf; derived test pairs shareable: unclear; commercial use: yes. No commercial-use clause (only no military use); clause 2.3.2 bars distributing derived data, so derived pairs stay inside the licensee and cannot go on a public leaderboard; faces must be disguised in published images.simulate MR/CT contrasts from the labels (e.g. BrainWeb-style), apply a known deformation; useful for head and neck multi-structure truth
MOBY / ROBY (4D mouse and rat phantoms)NURBS mouse (MOBY) and rat (ROBY) whole-body phantoms with cardiac and respiratory motion, ~1400 structuresunlimited synthetic: any motion phase, any anatomy variantorgan labels and motion vectors from the vectors program, by constructionboth imagesrequestrequest through the CVIT form; terms not shown publiclyaskneeds a request
the data needs a request or a data use agreement
paid or restricted licence likely, same family as XCAT Licence checked on 2026-10-10 at https://cvit.duke.edu/resource/moby-roby-phantoms/; derived test pairs shareable: unclear; commercial use: unclear. No licence text on the CVIT page.render CT or MR of the same MOBY at two breathing/cardiac phases, the vector output is the dense truth; realism gap to real micro-CT
MRI-DIRporcine head CT, T1, T2 in 4 compression states; synthetic head-neck MR before, during, afterphantom and 64 synthetic deformations10 gold markers (on CT only) and known synthetic deformationsboth imagesdirect: images and RTSTRUCT through NBIA, deformation fields through an IBM Aspera public link, no accountimages CC BY 4.0, deformation vector fields CC BY 3.0 (TCIA)openleft out
The nine 'subjects' are one porcine head phantom scanned by CT, T1 and T2 in four compression states (ten gold markers, visible on the CT only) and eight synthetic head and neck MR series (14 to 18 slices) made from one template patient, each with four synthetic deformations whose fields are on Aspera. One phantom and one template patient cannot give ten test and ten development pairs from different subjects, and the synthetic warps are a known-deformation truth the corpus makes itself.
MRiLab (MRI simulator)simulator (software): pairs are generated, not downloaded; GPU MRI simulation platform (Matlab)multi-contrast MR of one phantomalignment knownboth imagesdirectBSD-2-Clauseopenleft out
simulator software, no images to download (MRI), unmaintained since 2021.
not maintained since 2021simulate contrasts of one phantom and add a known warp
MRXCAT (cardiac MR simulation on XCAT)simulated cine and perfusion cardiac MR from XCAT anatomy with heart beat and breathingany number of frames/phases of one virtual patient; same anatomy at two cardiac or respiratory phasesXCAT organ masks per frame and XCAT motion vectors (vectors program) by constructionboth imagesdirectMRXCAT terms: use of original or modified software for non-commercial purposes; do not share derived code without permission; (c) University and ETH Zurichnon-commercialleft out
simulator software, no images to download; needs XCAT; non-commercial.
Matlab; example XCAT cine/perfusion files on the ETH page; brain-MR extension by Ayaz (github Aymen-Ayaz/Brain-MRI-Simulation, code CC0, data CC BY-NC-ND) Licence checked on 2026-10-10 at https://biomed.ee.ethz.ch/mrxcat.html; derived test pairs shareable: unclear; commercial use: no. Same text in the repo ReadMe.txt; download links by email; new anatomies also need a Duke XCAT licence.simulate the same XCAT anatomy at two cardiac or breathing phases, export XCAT motion vectors as the truth; realism gap to real CMR
MRXCAT 2.0 (biophysical LV + textured XCAT)synthetic cine CMR: biophysical left-ventricle model warped into an XCAT background with learned tissue texturecardiac phases of virtual patients (one example patient and one XCAT background shipped)LV displacement from the biophysical model (VTK meshes) and XCAT beating/breathing warps by constructionboth imagesdirectno licence in the repo; it bundles an adapted MRXCAT (non-commercial terms) and XCAT-generated tissue datanon-commercialleft out
simulator with one example patient; non-commercial terms bundled.
more realistic texture than MRXCAT 1; generating new backgrounds needs XCAT Licence checked on 2026-10-10 at https://github.com/sbuoso/mrxcat-2.0; derived test pairs shareable: unclear; commercial use: no. Commercial use is barred through the bundled MRXCAT terms; new backgrounds need XCAT (Duke licence).simulate end-diastole and end-systole of one virtual patient, dense LV displacement from the model is the truth; check texture warping does not leak motion cues
Multimodal ground truth datasets for abdominal registration (XCAT + CycleGAN, Bauer 2021)Synthetic abdominal CT, CBCT and T1 MR generated by CycleGAN from 56 XCAT phantoms, inhale and exhale; multimodal and respiratory, same virtual subject56 XCAT models x 3 modalities x 2 motion statesInherently co-registered images, XCAT organ masks and respiratory displacement fields, exact by constructionboth imagesdirectCC BY-NC 4.0 (heiDATA API)non-commercialleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): its licence is not shown to allow commercial use (CC BY-NC 4.0 (heiDATA API)). It was proposed as: Synthetic abdominal CT, CBCT and MR from XCAT; abdominal MR-CT and CT-CBCT already covered with real images and people-made truth; NC; 34 GB.
3D, synthetic texture; also usable as 2D slice pairs. Linked from Duke CVIT.
Multimodal MRI-HSI agar brain phantomsMRI vs hyperspectral/RGB/depth of agar brain phantoms13 phantomsFiducials, ArUco tags, tissue labels in MRI and HSI, tracked points (v2)both imagesdirectCC BY 4.0openwaits for a tool
hyperspectral and RGB images are 2D, against MRI of agar phantoms.
MUST / SIMUS / SIMUS3 (Matlab UltraSound Toolbox, PyMUST)simulator (software): pairs are generated, not downloaded; fast Fourier-domain US simulation for linear, convex and matrix arrays; speckle tracking2D/3D US of moving scatterersscatterer motion knownboth imagesdirectLGPL-3.0openleft out
simulator software, no images to download (ultrasound).
Python port PyMUSTdrive scatterers with a known cardiac or FE motion field
NCAT and MCAT (legacy Segars/Tsui phantoms)earlier NURBS (NCAT) and mathematical (MCAT) cardiac-torso phantoms with heart and breathing motionunlimited synthetic (historically)motion and labels by constructionboth imagesrequestno current distribution or licence page foundunknownneeds a request
the data needs a request or a data use agreement
status: superseded by XCAT; listed for completeness. PIXIE phantom: no source found Licence checked on 2026-10-10 at https://scholars.duke.edu/publication/1529895; derived test pairs shareable: unclear; commercial use: unclear. Superseded by XCAT; only the publication record exists.use XCAT or open alternatives instead
OpenCAS liver registration dataset (KIT)liver models deformed in silico and a silicone liver phantom indented under CT; undeformed to deformed3 in silico deformations + phantom experimentTeflon marker positions in CT before and after deformation; known in silico deformationboth imagesdirectnone stated; cite Suwelack et al. Med Phys 2014unknownleft outpriority 3
The archives hold meshes, STL surfaces and marker positions only: no image to register.
Meshes and surfaces mostly; archived site. Licence checked on 2026-10-10 at https://archive.iar.kit.edu/opencas/index.html%3Fq=node%252F25.html; derived test pairs shareable: unclear; commercial use: unclear. Archived KIT page only asks for a citation.
Patient-derived compressed digital breast phantoms (bCT, FE compression)821 breast phantoms segmented from patient dedicated breast CT, FE-compressed in CC view; uncompressed and compressed label volumes821 uncompressed/compressed pairs (labels: air, adipose, fibroglandular, skin; 0.273 mm)FE compression model gives the deformation; displacement fields not stated as included (unverified)both imagesdirectCC BY 4.0openleft out
label volumes only, no images; the FE displacements are not included.
label volumes, not images; huge compression (large deformation, thickness change)ask the authors for the FE nodal displacements or recompute with breastCompress; simulate bCT or DBT of both states to get image pairs with exact truth
Phantom FDAAnthropomorphic chest phantom with synthetic nodules, repeat CT under many protocols7 layouts x many repeatsKnown nodule size, density and premarked positionone imagedirectCC BY 3.0opentruth to make
no truth on both images; Repeats are identity and layouts differ in content: near-zero-motion checks only.
Fails: phantom not moved during repeats (identity transform); layouts differ in content.Repeats are identity and layouts differ in content: near-zero-motion checks only.
Plastimatch DRRsimulator (software): pairs are generated, not downloaded; plastimatch drr renders radiographs from CT; also deformable registration and warping toolsunlimited CT-to-DRR pairs; warp a CT with a known field then rendergeometry/pose known by constructionboth imagesdirectBSD-styleopenwaits for a tool
DRR software for 2D/3D pairs (waits for the 2D/3D tool).
simple ray-cast, fastwarp a CT by a known field, render DRRs of both: 2D/3D deformable truth
PLUS toolkit ultrasound simulatorsimulator (software): pairs are generated, not downloaded; simulated tracked US slices from surface models at tracked probe posestracked US slices vs CT/MR-derived modelsprobe pose known by constructionboth imagesdirectBSD-style (PLUS licence)openwaits for a tool
simulated tracked 2D US slices (2D/3D).
simple geometric US, no speckle physicssimulate tracked US sweeps of a CT-segmented model at known poses: US-CT pose truth
POSSUM (FSL MRI/fMRI simulator)simulator (software): pairs are generated, not downloaded; simulates MRI and fMRI from a segmented brain with rigid head motion, B0 effectsvolumes/time series of one brain under known motionmotion parameters knownboth imagesdirectFSL licence: free for non-commercial use; commercial use needs a licence from Oxford University Innovationnon-commercialleft out
simulator software, no images to download (MRI); FSL non-commercial licence.
commercial service would need a licencesimulate EPI with a prescribed motion trace; motion trace is the truth (includes distortion)
Prostate TRUS-probe FE motion simulations (UCL, Hu 2012 / Saeed 2020)NiftySim FE simulations of TRUS probe and balloon deforming the prostate, from patient MR segmentations100-500 simulations per patient; 160,000 on 320 patients (Saeed 2020)FE nodal displacementsboth imagesrequestnot released; no public pageunknownneeds a request
the data needs a request or a data use agreement
fails: not public; listed because it is the main prostate biomechanical source Licence checked on 2026-10-10: no primary page states terms; derived test pairs shareable: unclear; commercial use: unclear. No dataset record found; would have to be asked from UCL.ask UCL (Hu/Barratt) for a subset, or regenerate with NiftySim/FEBio on public prostate MR labels
QSM Reconstruction Challenge 2.0 head phantom (and 2016 data)digital head phantom from 7T MP2RAGEME maps; simulated multi-echo GRE magnitude/phase with ground-truth susceptibility and segmentation1 phantom (2 variants, 2 SNR levels); simulator can generate more contrastsknown susceptibility map and tissue segmentation, aligned by constructionboth imagesaccountRadboud University - Donders Institute - Human Data 1.0 (RU-DI-HD-1.0) data use agreement, restricted accessaskneeds an account
the download needs an account
Phantom is built from inherently co-registered quantitative maps; any simulated contrast is exactly aligned. Challenge page: https://qsm.snu.ac.kr/qsm-challenge. Licence checked on 2026-10-10 at https://api.datacite.org/dois/10.34973/m20r-jt17; derived test pairs shareable: unclear; commercial use: unclear. DataCite rights list; the DUA text page returned 404, so its commercial and sharing clauses were not read.simulate two contrasts (e.g. magnitude TE1 and susceptibility or R2*), warp one by a known field: exact multimodal truth with realistic anatomy
REVOLUS rotating-disk ultrasound (CREATIS)2D ultrasound of a rotating agar disk, simulated and in vitro (Verasonics), 1-5 rad/s~50-60 simulated sequences of 300 frames (pages disagree) + 5 in vitro sequences of 299 framesknown angular velocity gives the dense rotation field by constructionboth imagesdirectnone stated (page calls it an open-access database)unknownwaits for a tool
2D ultrasound of a rotating disk.
simple rigid rotation; good for speckle decorrelation tests only Licence checked on 2026-10-10 at https://humanheart-project.creatis.insa-lyon.fr/revolus.html; derived test pairs shareable: unclear; commercial use: unclear. No licence file in the REVOLUS Girder collection.
RIDER Phantom PET-CT (TCIA)repeat PET/CT of one NEMA-type phantom at many sites20 phantom scansknown phantom geometryboth imagesdirectTCIA (licence not seen)opentruth to make
rigid NEMA phantom repeat scans, truth to make from sphere centres; rigid and phantom-only.
Listed in Li 2023; repeat scans of a rigid phantom. Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Li 2023 dataset collection (ACM CSUR)).rigid phantom: sphere centres found by a program give exact rigid truth across sites
Simul@trophy (Khanal 2017) + example longitudinal imagesbiophysical atrophy model warps a real baseline MRI (OASIS) into follow-ups with prescribed volume lossany number of follow-ups per subject; 12 example images on NeuroVault collection 1392prescribed atrophy map and the model deformation field (fields not in the NeuroVault examples)both imagesdirectBSD-style, INRIA 2016, citation of two papers required; NeuroVault data CC0 by site policy (unverified)openleft out
simulator software; the 12 example images lack the deformation fields.
authors warn of circularity: intensity realism uses a registration fieldsimulate follow-ups from real baselines, keep the model field as truth; use the no-intensity-transfer mode to avoid circularity
SlicerCBM (UWA biomechanical brain-shift simulation)simulator (software): pairs are generated, not downloaded; meshless biomechanical brain model from preop MRI, outputs a predicted warped intraoperative MRIpreop MRI vs simulated intraop MRI with any boundary conditionmodel displacement field by constructionboth imagesdirectBSD-3-Clauseopenleft out
simulator software, no images to download (biomechanical brain shift).
3D Slicer extensionsimulate craniotomy brain shift on public preop MRIs, keep the field as truth; realism gap: no resection cavity or iUS appearance
Synthetic 4DCT(MRI) lung motion (Duetschler, PSI/ETH)4DMRI-derived lung motion fields used to animate a CT into synthetic 4DCT with known deformationmotion from 5 volunteers; 871 breathing cycles in the paperground-truth deformation vector fields by constructionboth imagesdirectCC BY 4.0 (meshes record 5016294 and code record 5010964)openneeds a request
lung meshes and a generator only; the CTs and generated 4DCTs are on request.
Users bring their own CT; the 13 patient CTs and generated 4DCTs are on request only. Truth inherits the generator's DIR. Licence checked on 2026-10-10 at https://zenodo.org/records/5016294; derived test pairs shareable: yes; commercial use: yes. Open part is the moving lung meshes and generator; the patient CTs and generated 4DCTs are on request, outside this licence.
TAM-ARa dynamic phantom CT/CBCT supine and uprightThoraco-abdominal phantom CT vs CBCT, supine and upright3 scansKnown phantom geometry (paper)directCC BY 4.0opentruth to make
three DICOM scans (CT supine, CBCT supine and upright) of a deformable phantom; no markers, contours or known displacement released
Two people mark 10 distinct phantom features on all three scans; known geometry bounds the error. Three scans make a demo, not a test set.
Thoracolumbar fascia ultrasound ground truth (gel pad)Dynamic US of gel-pad phantom moved at known speeds/distancesVideo setKnown displacements/velocitiesboth imagesdirectCC BY 4.0openwaits for a tool
ultrasound videos.
Thorax phantoms CT and Halcyon kV/MV CBCT with pacemakers (Kaggle)CT (Siemens SOMATOM) vs kV CBCT and MV CBCT (Varian Halcyon) of the same thorax phantom4 phantoms x 10 pacemakers, three modalities eachnone; pacemaker can and leads are rigid high-contrast markersaccountUnknownunknownneeds an account
the download needs an account
Kaggle download needs a login; licence not stated; uploader's paper unverifiedthreshold the pacemaker can corners and lead tips on each scan as landmarks, a person checks them; the phantom is rigid so a rigid fit is the truth
TIGRE (GPU cone-beam toolbox)simulator (software): pairs are generated, not downloaded; fast CBCT forward projection and reconstruction (FDK, iterative)CT vs simulated CBCT of the same volume, any geometrysame volume and known geometry (add a known warp for deformable truth)both imagesdirectBSD-3-Clauseopenleft out
simulator software, no images to download (CBCT).
used for 4D-CBCT simulation studiesproject a planning CT warped by a known field, reconstruct CBCT: CT-CBCT deformable truth with CBCT artefacts
UCSD deformable head and neck phantom with balloon tumour (Graves 2015)axial-slab head and neck phantom CT, balloon deflation simulates tumour shrinkagea few deformation statesnon-radiopaque surface markers measured outside the images; diode doseboth imageshiddennot releasedaskleft out
its images or its truth are not released
Not public.request from UCSD
UCSF patient-based simulated deformations and physical pelvic phantom (Nie 2013)patient CT (head and neck, prostate, craniospinal) with simulated bladder filling, mandible and vertebral moves; physical pelvic phantom3 patients, several deformations eachknown simulated DVFs from virtual deformation QA software, checked on a physical pelvic phantomboth imageshiddennot releasedaskleft out
its images or its truth are not released
Not public.known smooth warps on public CT, as for ImSimQA
UCSF pelvic deformable phantom with optical surface markers (Kirby 2013)axial-slab pelvic phantom CT, full vs empty bladder insert1non-radiopaque surface markers tracked by optical camera; rectum contoured on both imagesboth imageshiddennot releasedaskleft out
its images or its truth are not released
Markers invisible on CT, so the algorithms cannot use them. Not public.request from UCSF (Kirby, Pouliot)
UF/NCI computational human phantom library (NCI PHANTOM)pediatric and adult reference and size-dependent hybrid phantoms (height/weight grid); UF/MSK mesh library (357 adults) newermany between-subject pairs with matching organ setsorgan labels by constructionboth imagesrequestfree for non-commercial research via an NCI Software Transfer Agreement; commercial use needs a licence from the NCI Technology Transfer Centernon-commercialneeds a request
the data needs a request or a data use agreement
dosimetry phantoms, labels only Licence checked on 2026-10-10 at https://dceg.cancer.gov/tools/radiation-dosimetry-tools/phantoms-library; derived test pairs shareable: unclear; commercial use: no. Commercial licence is possible by asking NCI TTC (Dr. Kevin Chang).convert to CT numbers and deform with a known field; between-subject Dice on organs
Ultrasound phantom under controlled indentation (Borealis)2D B-mode and RF of an elastography phantom under robot-controlled indentation6 sequences, many stepsmeasured indentation and force per frame, no displacement fieldsdirectCC0 1.0opentruth to make
only indentation depth and force; label the inclusion on every frame (one structure in a phantom, apart) or build a finite-element model
inclusion boundary labelled in every frame; finite-element model of the phantom for a displacement field
Vendor-specific synthetic 2D speckle-tracking echo (Alessandrini 2018, KU Leuven)2D apical echo sequences from an E/M model with vendor-specific texture from 7 systems105 sequences (7 vendors x 3 views x 5 motion patterns)ground-truth tracking (seed points) and segmental strain from the modelboth imagesdirectCC BY-SA 4.0openwaits for a tool
2D echo sequences.
RF not shared (vendor NDA); scatter maps given so RF can be resimulated
Virtual Population ViP 3.x/4.x (IT'IS)whole-body anatomical models from volunteer MRI, >300 tissues, 0.5 mm; posing tool in Sim4Lifebetween-subject; posture changes of one model via the Poserlabels by construction; posture deformation known when posedboth imagesrequestpaid academic or commercial ViP licence; V3/V4 models only inside Sim4Life (encrypted .sab)askneeds a request
the data needs a request or a data use agreement
commercial; posing needs Sim4Life (commercial) Licence checked on 2026-10-10 at https://itis.swiss/virtual-population/virtual-population/overview-human-models/faq; derived test pairs shareable: no; commercial use: unclear. Commercial licence exists at a price; academic licence excludes commercially funded work; models cannot be exported from Sim4Life, so no image pairs to keep.pose one model in two postures and keep the Poser deformation as truth; simulate MR/CT from labels
XCAT / 4D XCAT phantomcomputational whole-body phantom with breathing and heart motion; render CT/MR/PET at any phaseunlimited syntheticknown motion vector fields and organ labels by constructionboth imagesrequestDuke licence via CVIT request or Duke OLV; terms not shown publiclyaskneeds a request
the data needs a request or a data use agreement
Unverified licence terms; synthetic. Licence checked on 2026-10-10 at https://cvit.duke.edu/resource/xcat-phantom-program/; derived test pairs shareable: unclear; commercial use: unclear. olv.duke.edu/xcat is behind an anti-bot wall; CVIT page only has a request button.
XCAT 3.0 personalized phantom library (Duke CVIT)over 2,500 CT-derived XCAT phantoms, up to 140 structures, voxel and mesh; chest phantoms with breathing motionbetween-subject pairs; same phantom at breathing phases (motion explorer shown)organ labels on every phantom (from 4 segmenters with QC); motion by constructionboth imagesrequestphantoms on request at cvit.duke.edu/resources; terms not shownaskneeds a request
the data needs a request or a data use agreement
labels come from automatic segmentation of real CT, so anatomy realistic but labels not expert Licence checked on 2026-10-10 at https://xcat-3.github.io/; derived test pairs shareable: unclear; commercial use: unclear. Project page and CVIT page only point to the request form.render two breathing states with DukeSim (or any CT simulator) and keep XCAT motion vectors; or between-subject Dice
XCIST (open CatSim CT simulator)simulator (software): pairs are generated, not downloaded; CT projection simulation (spectra, detectors, scatter) and reconstruction from voxel/analytic phantomstwo scans of any phantom statetransform between phantom states known by constructionboth imagesdirectBSD-3-Clauseopenleft out
simulator software, no images to download (CT).
realistic CT physics; phantoms bring their own anatomysimulate CT of a phantom (e.g. XCAT or a labelled patient CT) before and after a known warp; realism gap in texture

2D · 214

DatasetProblemPairsChecked againstAccessLicenceStatusNotesTo make the truth
10x Xenium FFPE human breast preview (Janesick 2023)FFPE breast cancer: Xenium on serial sections (replicates), post-Xenium H&E and IF on the same section, Visium CytAssist on an adjacent section; same section and serial sectionsSame-section DAPI-H&E pairs; serial-section pairs (unverified counts)10x alignment of H&E to Xenium (keypoints/affine, unverified)directnone on the preview page; raw data also in GEO GSE243280unknownleft out
one or two sections: no ten independent pairs; its H&E-to-Xenium truth is 10x's own alignment
Details from memory of the paper; page not reachable during this sweep. Licence checked on 2026-10-10 at https://pmc.ncbi.nlm.nih.gov/articles/PMC10730913/; derived test pairs shareable: unclear; commercial use: unclear. Live 10x page is behind a bot wall; archived preview page shows no licence. Sister 10x Xenium breast dataset pages say Creative Commons Attribution, so CC BY is likely but not stated for this one.Same-section DAPI vs H&E: exact truth under a known warp; serial sections: transcript-defined cell types as labels on both
3DBodyTex.DermSynth3D2D renders of 3D body scans with synthetic skin lesions blended in (Fitzpatrick17k lesions), many views of the same body100K rendered images from 215 scanslesion, skin and body-part masks per render; renders from known cameras of one mesh, so exact cross-view correspondence is computableboth imagesrequestlicence agreement signed by the recipient and the institution's research administration director; citation requiredaskneeds a request
the data needs a request or a data use agreement
Synthetic lesions; exact multi-view truth by construction if camera poses are shipped (check) Licence checked on 2026-10-10 at https://cvi2.uni.lu/3dbodytexdermsynth/; derived test pairs shareable: unclear; commercial use: unclear. Agreement terms are not published; needed because of the volunteer models' consent agreements.render two views of the same mesh with the DermSynth3D code: exact dense correspondence from the mesh
AAOF Craniofacial Growth Legacy CollectionLateral and frontal (PA) cephalograms of untreated children followed through growth (Bolton-Brush, Burlington, Denver, Fels, Forsyth, Iowa, Mathews, Michigan, Oregon); same subject over yearsAbout 12,000 lateral and 8,500 PA cephalograms from 941 subjects at several agesCephalometric measurements for a subset (10 subjects per collection); no landmark coordinates publishedone imagerequestfree online use and download for the orthodontic community, education and research; others must ask permission from the collection of originaskneeds a request
the data needs a request or a data use agreement
The best open source of temporal cephalogram pairs. Digitised films of varying quality. Licence checked on 2026-10-10 at https://www.aaoflegacycollection.org/aaof_home.html; derived test pairs shareable: unclear; commercial use: unclear. High-resolution images from the source collections, sometimes for a fee; Burlington needs an extra form; cite AAOF and the collection.Two people place the 19 ISBI-2015 landmarks on 50 subjects at two ages; anterior cranial base (sella, nasion, cribriform plate) is the classic stable structure for superimposition
AAPM DL-Spectral CT Grand Challenge 2022 datasimulated 2D breast CT: low- and high-kVp sinograms and FBP images with ground-truth tissue maps (adipose, fibroglandular, calcification)1,000 training casesexact tissue maps from the simulationboth imagesdirectCC BY 4.0openleft out
simulated 2D breast CT phantoms; a known warp on simulated textures adds nothing the known-warp sets lack
Synthetic; all images per case aligned by construction.warp the high-kVp image or a tissue map by a known field and register the low-kVp image: exact 2D multimodal truth with exact labels
AASCE 2019 / SpineWeb Dataset 16AP spinal radiographs of scoliosis patients (London Health Sciences Centre); between subjects609 images (481 train, 128 test) + 98 hidden test68 landmarks (4 corners of 17 vertebrae T1-L5) by two doctors; Kordon's curated corrections on Zenodoboth imagesdirectno licence; SpineWeb asks any publication to cite Wu et al. MICCAI 2017; Zenodo annotation record has no licenceunknownneeds a request
68 corner points by two doctors would pass, but only the annotations are on Zenodo: the SpineWeb image host no longer resolves; ask the authors for the images
Status change risk: spineweb.digitalimaginggroup.ca does not resolve (2026-10-10); images may need a mirror. status change: Round 4 count: listed in 2 of the 38 round-4 surveys and benchmark papers (Zou 2022 review (Front Oncol); Hammoudeh 2023 survey). Licence checked on 2026-10-10 at http://web.archive.org/web/20240613042556/http://spineweb.digitalimaginggroup.ca/index.php?n=Main.Datasets; derived test pairs shareable: unclear; commercial use: unclear. SpineWeb host is down; other SpineWeb sets are explicitly non-commercial, Dataset 16 says nothing beyond the citation.
ADMANI (BreastScreen Victoria, BRAIx)screening mammograms (4 views) of the same women across biennial rounds4.4 million images, 629,863 women, 1.05 million episodes (longitudinal ADMANI2 and 3)radiologist annotations of cancers at reading; histopathology outcomesone imagehiddenunder licence agreement with the state screening programme; future availability being developed; a 40,000-image subset went to the RSNA 2022 challengeaskleft out
its images or its truth are not released
the RSNA 2022 challenge subset (already listed) came from here; full longitudinal set not availableprior/current pairs: nipple, pectoral line and calcification clusters as landmarks by two readers
AGE challenge (AS-OCT angle closure)Anterior-segment OCT B-scans of the angle, inter-subject (and left/right angle within a scan)4,800 images, 199 patientsScleral spur coordinates (both angles) by expertsboth imagesaccountno data licence; grand-challenge terms of participation for registered teams (true identity, citation)askneeds an account
the download needs an account
Only 2 landmarks per image: weak for registration; inter-subject AS-OCT. Licence checked on 2026-10-10 at https://age.grand-challenge.org/Rules/; derived test pairs shareable: unclear; commercial use: unclear. Download page is gated; rules say nothing about reuse of the data after the challenge.Add iris root, corneal apex and lens points by two people on 50 images.
AI-READI (multi-device retinal imaging)Same participant imaged on several retinal devices in one visit (colour fundus, IR, FAF, OCT, OCTA; devices such as Eidon, Optomed, Spectralis, Maestro2, Triton, Cirrus - from memory)up to 2,280 participants (v3); per-modality counts unverifiedNone for registrationone imageaccountcustom AI-READI Data License Agreement (public set) or DTUA via the Data Access Committee (restricted set or non-diabetes use)askneeds an account
the download needs an account
Device list not confirmed on the pages read. Cross-device same-visit pairs are rare elsewhere. Licence checked on 2026-10-10 at https://docs.aireadi.org/docs/3/preliminary/license; derived test pairs shareable: no; commercial use: unclear. DLA allows commercial or non-commercial research but only on type 2 diabetes; a registration benchmark needs a DTUA. Storage only on institution servers or clouds with a HIPAA BAA; no sharing outside the institution.Same-device OCT/IR pairs aligned by construction give known-warp truth; cross-device: two people place 8 bifurcation pairs on 30 eyes.
Alex iCare longitudinal OCT (Elkholy 2026)clinical OCT B-scans of normal, DME and AMD eyes in visit folders, same patient over time42,061 scans; visit sub-folders 3 months apart (the text also says one visit per patient: inconsistent)nonerequestavailable after agreement with iCare centre and Alamein International Universityaskneeds a request
the data needs a request or a data use agreement
B-scans resized and standardised, so geometry may be lost; longitudinal claim inconsistent in the papersame-location B-scans across visits: retinal layer boundaries from a public segmenter as a weak check
AngioLens (pre and post treatment coronary angiograms)Invasive coronary angiograms before and after PCI from one hospital (Souad Kafafi)3,416 pre-treatment and 3,598 post-treatment images; same-patient matching unverifiedClass folders onlyone imagedirectCC BY 4.0openneeds a request
the Zenodo record is restricted (no file is public); no spatial truth either
Fails truth; record files not visible via APIPair pre/post views of the same patient and projection angle; two people mark 10 bifurcations on each
AngioReport (APTOS 2023 FA and ICGA)Fluorescein angiography and ICG angiography frame series per eye (few to hundreds of frames), Rajavithi Hospital55,361 images, 3,179 eyes (81.8% both FA and ICGA)Report-style labels (24 conditions, lesion type/location); no correspondencesone imagerequestcredentialed access by application (reviewed in 5-10 workdays); no licence namedaskneeds a request
the data needs a request or a data use agreement
Simultaneous FA/ICGA composites (Heidelberg) are side-by-side, aligned by construction. Fails truth. Licence checked on 2026-10-10 at https://bjo.bmj.com/content/109/11/1283; derived test pairs shareable: unclear; commercial use: unclear. Data availability statement points to https://asiateleophth.org/cross-country-datasets/ (APTOS); cite Xu et al. BJO 2025.Simultaneous FA/ICGA frames are aligned by construction: apply a known warp to one for exact multimodal truth; early-to-late frames: tracked bifurcations verified by a person.
ANHIRhistology whole-slide images, 18 stains, same tissue481 (230 with public landmarks)landmarks; the 251 evaluation pairs are hiddenboth imagesaccountCC BY-NC-SAnon-commercialplannedChosen for the 2D track. status change: Round 4 count: listed in 5 of the 38 round-4 surveys and benchmark papers (Chen 2025 survey (MedIA); WSI registration review 2025; Li 2023 dataset collection (ACM CSUR); ACROBAT 2023; ANHIR/BIRL 2020). Used in 6 of the 2D and microscopy registration papers read (HistoReg (Greedy) 2019; VALIS 2023 (per review); SFG-CNN Ge (per review); DeeperHistReg 2024; ...).
AnnoCervcolposcopy images per case: acetic acid, iodine and green-filtered views of the same cervix527 images from 100 recordsexpert segmentations of healthy and pathological tissue on acetic acid images onlyone imagedirectCC BY 4.0opentruth to make
acetic-acid and iodine views of 100 cervices (CC BY 4.0) with masks on the acetic image only; two people click 8 landmarks on 30 pairs
Masks only on acetic images, so no cross-stage label check Licence checked on 2026-10-10 at https://github.com/iclx/AnnoCerv; derived test pairs shareable: yes; commercial use: yes. README licence section and GitHub licence field both CC BY 4.0; images are the authors' own colposcopy records.two people click 8 landmarks on 30 aceto-iodine pairs
AO retinal longitudinal montage pairs with manual alignment (Penn, Chen et al. 2019)adaptive optics SLO triplets (confocal, split detection, dark field), same retinal location at two visitstest: 20 locations in 4 subjects with manual alignment; development: 30 locations in 5 subjects (figshare 12003651); montage overlap pairs (figshare 3717249)manual alignment of each test pair by the authorsboth imagesdirectCC BY 4.0opentruth to make
the 'manual alignments' are rendered TIFF montages and the cone positions a detector's; 4 subjects: match cones by program, a person checks, and count only non-overlapping locations
found through the figshare API; cone-scale 2D, very different from fundus setsa second person marks 20 matching cones per pair on 10 test pairs to check the manual alignment
ARCADE (coronary X-ray angiography segmentation and stenosis)Single coronary angiography frames (up to 12 per patient, 6 views); between subjects3,000 frames (1,500 SYNTAX segments, 1,500 stenosis)Expert masks of 25-26 SYNTAX coronary segments and stenosis regionsboth imagesdirectCC0 1.0 (HF mirror card; Zenodo not re-checked)openleft out
single frames of different patients and views: no pair to register
Fails for temporal pairs (single frames); named segments allow same-view inter-subject overlap only.None practical; only usable for synthetic warps of single frames
AREDS fundus photographs (dbGaP phs000001)Macular-field (F2) colour fundus photographs of AMD study eyes over years of follow-up, same eye134,500 F2 photos, 4,613 participants (~29 per participant, implying many follow-up pairs)AMD severity grades by reading centre; no correspondencesone imagecontrolleddbGaP Data Use Certificationaskneeds a request
the data needs a request or a data use agreement
Visit structure inferred from counts; 'Ocular Photographs' inventory not read. Fails truth.Two people place 10 bifurcation pairs on 30 longitudinal pairs; drusen/GA grades as weak change check.
AREDS2 and METforMIN longitudinal FAF with geographic atrophy contoursfundus autofluorescence of eyes with geographic atrophy at several visits (AREDS2: 830 images, 257 eyes; METforMIN: 273 images, 71 eyes)consecutive-visit pairs per eye (hundreds)GA contours per visit by reading-centre graders; authors registered visits and fixed failures with manual vessel bifurcationsboth imagescontrolledunverified (AREDS2 via NEI/dbGaP request; METforMIN via investigators)askneeds a request
the data needs a request or a data use agreement
GA grows between visits, so contours are not a pure registration truth; vessel bifurcations arevessel bifurcations found by a program and verified by a person on 20 pairs; GA contours only outside growth area
BCDR (Breast Cancer Digital Repository, film and digital)film (BCDR-FM) and full-field digital (BCDR-DM) mammograms, CC and MLO, Porto; some patients with more than one studyFM: 1,010 patients, 1,125 studies; DM: 724 patients, 1,042 studies (so some repeat studies); CC/MLO of each breastlesion outlines by radiologists on MLO and/or CC views (1,517 FM and 818 DM segmentations)both imagesaccountfree for registered users (terms unverified)unknownneeds an account
the download needs an account
lesion outlines on both views give CC-MLO correspondence; prior/current pairs exist only where a patient has several studies (count unverified)for repeat studies: same lesion outlined on both studies gives lesion-level truth; add nipple and pectoral landmarks by two people
BCN20000dermoscopy, several images per lesion with capture dates, Hospital Clinic Barcelona18,946 images of 5,583 lesions (so many repeats); capture date givennone spatialone imagedirectCC BY 4.0opentruth to make
repeat dermoscopy of one lesion with dates (CC BY 4.0), no spatial truth: two people click 6-10 structures on 30 pairs
Open licence and dates: repeat images of the same lesion, some on different daystwo people click 6-10 structure landmarks on 30 same-lesion pairs
BIMCV-COVID19+Chest radiographs (CR, DX) and CT of COVID-19 patients, Valencia region, several sessions per patient21,342 CR, 34,829 DX and 7,918 CT studies (all iterations)UMLS findings with anatomical labels; 23 images semantically segmented by radiologistsone imagerequestBIMCV research use agreement (no named licence)non-commercialneeds a request
the data needs a request or a data use agreement
Sessions per patient not stated on the page (BIDS/MIDS layout with sessions, unverified). Some patients have CXR and CT close in time.Pair sessions per patient; landmarks by two people on 30 pairs; CT-derived DRRs of the same patient could give a 2D/3D check
BRAXChest radiographs, Hospital Israelita Albert Einstein (Brazil), some patients with several studies40,967 images, 24,959 studies, 19,351 patients14 report-derived labels; nothing spatialone imagecontrolledPhysioNet Credentialed Health Data License 1.5.0askneeds a request
the data needs a request or a data use agreement
Few repeats per patient (about 1.3 studies per patient).Pair repeat studies; two people place 20 landmarks on 30 pairs
Brazilian dental panoramic dataset (EHAML age estimation)Panoramic radiographs, University of Pernambuco; some patients with several panoramics over time12,827 images from 10,036 patients (about 2,800 repeat images)Age onlyone imagerequestcustom terms: research use only; redistribution, modification and commercial use illegal without the authors' permission; no third-party sharingnon-commercialneeds a request
the data needs a request or a data use agreement
One of very few panoramic sets with repeat exams of the same person; dates unverified. Licence checked on 2026-10-10 at https://github.com/willianfco/ehaml-brazilians-dataset; derived test pairs shareable: unclear; commercial use: no. Access by e-mail from an institutional account with the agreement text; derived pairs count as modification, so internal copies need the authors' permission.Pick 30 repeat pairs; tooth masks from an O2PR model checked by a dentist, plus 20 landmarks (cusps, root apices, mental foramina) by two people
BrixIA COVID-19 datasetChest radiographs of COVID-19 inpatients, Brescia (Italy), many with repeat radiographs4,703 CXRs (patients and repeats per patient unverified)Brixia score (6 lung regions, 0-3) by radiologists; no landmarksone imagerequestResearch use only, conditions on the dataset websitenon-commercialneeds a request
the data needs a request or a data use agreement
Region scores per lung zone give a coarse per-region change label between visits.Two people place 20 landmarks on 30 repeat pairs; region scores locate expected intensity change
BS-80K (whole-body bone scans)Tc-99m bone scintigraphy, anterior and posterior whole-body views; anterior-to-posterior (mirrored) within patient, or between patients3,247 patients, ANT+POST each (82,544 images incl. 13 regional crops)Hot-spot bounding boxes by nuclear medicine physicians on each view; anatomical region segmentations released later via 2DAnatomyDatasets (Seibold)both imagesdirectnone stated (Google Drive folder from the paper; paper calls it open-access)unknowntruth to make
anterior and posterior bone scans of one patient (mirrored) with hot-spot boxes not linked across views; two people click 15 skeletal landmarks on 20 patients; no licence stated
No follow-up pairs. ANT/POST boxes are not linked as correspondences. Link is a Google Drive folder given in the paper (Comput Biol Med 2022). Licence checked on 2026-10-10 at https://drive.google.com/drive/folders/1DOBkLXgQeREQjF-nQIGNBBzPCb5s7RNu?usp=sharing; derived test pairs shareable: unclear; commercial use: unclear. No licence file or terms found; no official repository.Region masks on both views give a Dice check for ANT-to-mirrored-POST and inter-subject warps; for a TRE, two people click 15 skeletal landmarks (joints, iliac crests) on 20 patients.
BSE-JSRT and ML-BoneSuppression pairsChest radiograph and its bone-suppressed version (JSRT images processed by Budapest University of Technology), aligned by construction247 JSRT pairs; IEEE DataPort set: 4,080 source/target pairs (augmented copies)Same geometry by construction (pixel-to-pixel)both imagesaccountIEEE DataPort record says CC BY 4.0, but the radiographs are JSRT images, whose owner forbids distribution and commercial useaskneeds an account
the download needs an account
Aligned by construction, so any known warp of one image gives exact multimodal-like truth (bone vs soft-tissue appearance). Licence checked on 2026-10-10 at https://api.datacite.org/dois/10.21227/xnb5-hg35; derived test pairs shareable: unclear; commercial use: unclear. Third-party CC BY on top of JSRT images is doubtful; JSRT terms at http://db.jsrt.or.jp/eng-01.php govern the base images.Apply a known smooth deformation to the soft-tissue image and register it back to the original
BUU-LSPINE (Burapha lumbar spine)Lumbar spine radiographs, AP and lateral views of the same patient, Burapha University Hospital; between subjects3,600 patients, 7,200 images (paper); portal v0.1 lists 400 AP/LA pairsCorner points of L1-L5 (and S1 upper edge on lateral), spondylolisthesis and LSTV labels by physiciansboth imagesdirectBurapha Faculty of Informatics end-user licence agreement: no re-identification, acknowledge and cite, no endorsement; full set by Google Form requestaskin the corpuspriority 3
In the corpus: lateral views of the 2,000-patient release paired between patients on the 22 endplate corners, 30 test and 30 development pairs; EULA without terms on redistribution (ask).
Pixel spacing given. Portal size differs from paper. Licence checked on 2026-10-10 at https://services.informatics.buu.ac.th/spine/; derived test pairs shareable: unclear; commercial use: yes. EULA puts no limit on commercial use (it even mentions the end-user's products) and says nothing on redistribution.
CBIS-DDSM (CC/MLO lesion correspondences)Digitised film mammograms; same lesion on CC and MLO views of the same breast1,566 participants, 10,239 images; 1,645 abnormality casesROI masks of each abnormality, curated by radiologists, on the views where it is seenboth imagesdirectCC BY 3.0openleft out
CC and MLO are different projections: lesion matching, not image registration; no priors
CC and MLO are different projections: no dense registration, only lesion-to-lesion matching (ipsilateral correspondence). No priors. status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (MGRegBench 2025).Use lesion centroids on CC and MLO as a correspondence test (distance-to-nipple consistency)
CDD-CESM (contrast-enhanced spectral mammography)Low-energy and recombined (subtracted) contrast mammograms of the same breast and view, Cairo326 patients, 2,006 images (about 1,000 low-energy/subtracted pairs if every view has both)Same acquisition (dual-energy), aligned by construction except small motion; hand-drawn lesion segmentationsboth imagesdirectCC BY 4.0opentruth to make
low-energy and recombined images of one exposure are aligned by construction: a known warp gives an exact cross-appearance truth (TCIA, CC BY 4.0)
Gives exact multimodal pairs for a known-warp benchmark (anatomy vs contrast uptake appearance). status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (MGRegBench 2025).Apply a known smooth deformation to the subtracted image and register it to the low-energy image
Cervical spine flexion-neutral-extension radiographs with 22 landmarks (Yan 2022)lateral cervical radiographs in flexion, neutral and extension, same patient932 patients x 3 views (2,796 images)22 vertebral corner landmarks (C2-C7) per view; test set of 100 patients by three radiologistsboth imageshiddenin-house (hospital PACS), not releasedaskleft out
its images or its truth are not released
rigid-per-vertebra registration pairs with landmarks; also Medical Metrics lumbar flexion-extension sets (161 and 415 volunteers, QMA-tracked) are proprietary
CGMH Pelvis X-ray segmentation databaseAP pelvis radiographs, Chang Gung Memorial Hospital; between subjects400 imagesHip segmentation made with the BoneFinder tool (automatic shape model, check unverified)both imagesdirectMIT (repository licence; images and labels are in the repo)opentruth to make
hip contours from BoneFinder, a shape-model program; an expert checks them on the test images first
BoneFinder points underlying the masks are point-correspondent if exported (unverified). Licence checked on 2026-10-10 at https://github.com/yaufan/Pelvis-X-ray_Segmentation_Database; derived test pairs shareable: yes; commercial use: yes. GitHub now detects an MIT licence for the repo; MIT is a software licence applied to the data files.Expert check of BoneFinder contours on 50 images
CHECK (Cohort Hip & Cohort Knee) radiographsHip and knee radiographs of 1,002 Dutch adults with early OA, baseline, 2, 5, 8 and 10 yearsUp to 5 visits per participantKL grades per visit; no landmarksone imagerequestCC BY 4.0 (DANS dataset, open access)openneeds a request
the data needs a request or a data use agreement
Records disagree on whether images are included; confirm with DANS. Licence checked on 2026-10-10 at https://api.datacite.org/dois/10.17026/DANS-XEX-HZWW; derived test pairs shareable: yes; commercial use: yes. DataCite record says X-rays of hips and knees at baseline, 2 and 5 years are in the dataset (T0..zip files, ~9.6 GB each); DANS page is behind an anti-bot wall, so file-level access was not checked.Two people place hip/knee landmarks (femoral head centre, teardrop, tibial spines) on 30 participants across visits
Chest ImaGenome (gold standard pairs)Frontal chest radiographs from MIMIC-CXR, same patient over timeGold set: 500 patients with 2+ exams, 1,000 images; silver: 242,072 images, >670,000 comparison relationsBounding boxes for 29 anatomical regions (lungs, zones, hila, heart, mediastinum, costophrenic angles, clavicles...), automatic, manually validated or corrected on the 1,000 gold images; comparison relations (improved/worsened/no change) per regionboth imagescontrolledPhysioNet Credentialed Health Data License 1.5.0askneeds a request
the data needs a request or a data use agreement
Box centres/corners per anatomy on both images of a pair give a coarse registration check (box IoU or centre error). GRCD and TRACE re-derive cleaned pair lists from it.Use gold-set box corners as coarse landmarks; add 15 point landmarks per pair on 100 gold pairs by two readers
Chest X-ray landmark dataset (Gaggion)Chest radiographs from JSRT, Padchest, Montgomery, Shenzhen; inter-subject911 imagesDense contour landmarks with point correspondence (lungs 44+50, heart 26), fitted by a HybridGNet ensemble to manual masksboth imagesdirectno licence for the annotations; images must be fetched from JSRT, Shenzhen, Montgomery and PadChest under their own termsunknownleft out
its points are a network ensemble's fit to masks, with heart points for JSRT and Padchest only; images from four other sets under their terms
Gives landmark error for 2D inter-subject registration; images from source sites. Licence checked on 2026-10-10 at https://github.com/ngaggion/Chest-xray-landmark-dataset; derived test pairs shareable: unclear; commercial use: unclear. README only asks to cite the TMI 2022 and MICCAI 2021 papers; PadChest itself is non-commercial.
CheXmaskChest radiographs from 5 large databases; inter-subject657,566 masksAutomatic HybridGNet lung/heart contours with RCA quality scoresboth imagesdirectCC BY 4.0 (masks); source images under their own termsopenleft out
network masks without images of its own; kept as the automatic check for NIH ChestX-ray14 follow-up pairs
Fails expert truth.
CheXpert / CheXpert Plus (repeat studies)Chest radiographs, Stanford Hospital 2002-2017, same patient over timeCheXpert: 224,316 images of 65,240 patients; CheXpert Plus: 187,711 studies of 64,725 patients, with reportsNone spatial for pairs (CheXlocalize pathology masks on single validation/test images only)one imageaccountStanford AIMI research use agreement (licence text not shown on the page)non-commercialneeds an account
the download needs an account
Study order recoverable from study numbers per patient. CheXmask covers CheXpert images. status change: Round 4 count: listed in 2 of the 38 round-4 surveys and benchmark papers (Azad 2023 ViT review; Foundation models review 2026).Pick PA pairs per patient; two people place 20 landmarks on 50 pairs; CheXmask lung/heart masks as a weak overlap check
CheXTemporalPrior and current frontal chest radiographs, same patient, from CheXpert, MIMIC-CXR and ReXGradientGold: 1,787 pair-finding examples; silver: 282,214 pair-finding examples (128,071 studies)Gold: 4,702 radiologist bounding boxes with correspondence IDs linking matched findings across the two images (single radiologist per case); silver: MedGemma labels and 168,140 automatic anatomy masksboth imagesdirectCC BY-NC 4.0 (annotations); images under each parent dataset's termsnon-commercialneeds an account
matched radiologist boxes on findings that change between visits, annotations only: the images come from MIMIC-CXR (PhysioNet credentialing), CheXpert and ReXGradient, all behind accounts
Anonymous HF account (under review, May 2026); host may move. Matched boxes are on pathology, which changes between visits, so box centre error is a weak check.Use matched box pairs for stable findings only; add anatomical landmarks on 50 gold pairs by two readers; silver anatomy masks as an overlap check
Children's dental panoramic radiographs (Zhang 2023)Panoramic radiographs of children aged 2-13 (Hangzhou) plus re-annotated adult sets; between subjects193 child images (100 annotated) + 2,692 adult images with tooth segmentationWhole-dentition tooth masks and caries/disease boxes by dentistsboth imagesdirectCC0opentruth to make
one binary mask for all teeth; a dentist labels teeth one by one on 30 images
Masks are binary (all teeth), not per-tooth: weak inter-subject truth. Licence checked on 2026-10-10 at https://figshare.com/articles/dataset/21621705; derived test pairs shareable: yes; commercial use: yes. Dataset and supplement articles of collection 10.6084/m9.figshare.c.6317013 are both CC0.Per-tooth labels by a dentist on 30 images
CIMA (ANHIR subset)histology: lung lesion, lobe, mammary gland108landmarks on GitHub for all pairsboth imagesdirectCC BY-NC-SA (ANHIR data usage agreement); landmark repo BSD-3-Clausenon-commercialleft out
its images are part of ANHIR (planned, owner downloads); Borda's repository holds some of these landmarks twice (annotators PS and JB), useful later for the truth's own error
Image host did not answer; overlap with ANHIR evaluation pairs unverified. Licence checked on 2026-10-10 at https://anhir.grand-challenge.org/Data/; derived test pairs shareable: yes; commercial use: no. CIMA lesions, lung-lobes and mammary-gland images are part of ANHIR; original CIMA host now 404. ShareAlike on derived pairs.
CL-Detection 2023Lateral cephalograms from 3 centres (ISBI 2015 + PKU + Shenzhen); inter-subject600 images (150 test hidden)38 landmarks by cliniciansboth imagesaccountCC BY-NC-NDnon-commercialneeds an account
the download needs an account
Use ISBI/PKU parts directly to avoid ND/NC.
CMMD (Chinese Mammography Database)Mammograms (both breasts) of Chinese patients with biopsy-confirmed lesions1,775 patients, 5,202 imagesBenign/malignant and molecular subtype labels; no ROIsone imagedirectCC BY 4.0opentruth to make
no priors and no spatial truth; left and mirrored right breast of one woman with reader-placed points
Fails: no priors found, no spatial truth. status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (MGRegBench 2025).Contralateral CC/MLO pairs with reader-placed nipple and pectoral points
CO2Wounds-V2 (leprosy chronic wounds)smartphone photos of leprosy ulcers during treatment sessions, 96 patients over 21 months764 images (607 labelled); same-patient repeats not identified in the release (unverified)wound masks (COCO and PNG)one imagedirectCC BY-NC 3.0non-commercialleft out
visits of one patient are not linked in the release: no pair to form
Same clinic as LUTSeg; patient linkage likely missing Licence checked on 2026-10-10 at https://data.mendeley.com/datasets/s2w7rjwz49/2; derived test pairs shareable: yes; commercial use: no. Mendeley Data licence field.only by construction (known warp) unless patient IDs are obtained
CORIS glaucoma optic-nerve pairs (EyeLiner)Longitudinal optic-nerve-head fundus pairs (Nidek AFC-330), glaucoma, same eye55 pairs10 clinician keypoints per pairboth imageshiddeninternal (Univ. Colorado)unknownleft out
its images or its truth are not released
Private.
Corneal sub-basal nerve plexus mosaics in Parkinson's disease (Badian 2021)Corneal confocal microscopy mosaics (avg 48 fields each), 42 PD and 13 controlsmosaics onlyClinical data and nerve metrics; no correspondencesone imagedirectCC BY 4.0openleft out
finished mosaics only, no second image: only a self-warp, which is not the protocol's known-warp case
Sci Data 2021. Fails pair criterion. Licence checked on 2026-10-10 at https://figshare.com/articles/dataset/Wide-field_mosaic_dataset_of_the_corneal_subbasal_nerve_plexus_in_Parkinson_s_disease_using_in_vivo_confocal_microscopy/14481249; derived test pairs shareable: yes; commercial use: yes. Licence read from the figshare API record.Synthetic tile pairs cut from mosaics with known shift and warp.
Corneal sub-basal nerve plexus mosaics, healthy and type 2 diabetes (Lagali 2018)Laser-scanning corneal confocal microscopy mosaics of the sub-basal nerve plexus, both eyes322 mosaic files; several mosaics per eye (overlap unverified)Automatic nerve tracing and density; no correspondencesone imagedirectCC BY 4.0openleft out
finished mosaics only, raw frames not included: no real pair
Sci Data 2018. Raw frames not included. Fails pair criterion unless mosaics of one eye overlap.Crop overlapping tiles from a mosaic with a known shift plus smooth warp; nerve tracings warped as truth.
COVID-19 Image Data Collection (ieee8023)Chest radiographs (and some CT) of COVID-19 and other pneumonia from papers and websites, with day offsets for follow-upsAbout 900 images (unverified), a minority of patients with follow-up imagesLung boxes and segmentations (General Blockchain, v7labs, 517 images), lungVAE masks, severity scores; masks on many imagesboth imagesdirectPer image (Apache 2.0, CC BY-NC-SA 4.0, CC BY 4.0); metadata CC BY-NC-SA 4.0non-commercialleft out
images copied from publications (cropped, rescaled JPEG), few follow-ups, mixed licences; NIH ChestX-ray14 and COVID-19-NY-SBU give clean repeat radiographs
Mixed quality (many images from publications, cropped, JPEG). Repo inactive since 2022.Pairs from same patient with different offsets; v7labs lung masks as weak check; add 15 landmarks on the best 20 pairs
COVID-19-ARChest radiographs and CT of COVID-19 patients (Arkansas), some with repeat studies105 patients, 256 studies, 31,935 images (CT, CR, DX)Clinical spreadsheet with key radiology findings; no image annotationsone imagedirectCC BY 4.0opentruth to make
a few repeat chest radiographs a patient, no image annotation; two people place landmarks on the repeat pairs
Small; longitudinal nature not stated on the page.Landmarks on the few CXR repeat pairs; CXR-CT of same patient for 2D/3D
COVID-19-NY-SBU (TCIA)Chest radiographs plus CT/MR/PET of COVID-19 inpatients, one encounter per patient, repeated imaging during the stay1,384 patients, 7,361 studies, 562,376 imagesClinical data (OMOP); no image annotationsone imagedirectCC BY 4.0opentruth to make
frequent repeat radiographs within days, open licence, no annotation; two people place 20 landmarks on 50 pairs
Open licence and frequent repeat CXRs within days: good source of short-interval pairs with tubes and lines moving.Pick 50 day-apart CXR pairs; two people place 20 landmarks (incl. tube/line tips, which move); lung masks from a public segmenter as weak check
CSAW-CC (mammography)Swedish screening mammograms with all repeated screenings per woman (case-control, 2008-2015)873 cancer cases and 7,850 controls public; one study kept 406 cases and 6,053 controls with 2+ examsCancer outcome; tumour pixel annotations for cancer cases (unverified)one imagerequestCC BY 4.0 (Swedish National Data Service record); image files restricted, sent on request because of their size (~2.5 TB)openneeds a request
the data needs a request or a data use agreement
Regular screening intervals (12-36 months in one study). status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (MGRegBench 2025). Licence checked on 2026-10-10 at https://researchdata.se/en/catalogue/dataset/2021-204-1; derived test pairs shareable: yes; commercial use: yes. Licence field on the record is CC BY 4.0; the access request may still add conditions, which the record does not show.Two readers mark 10 stable points (nipple, vessels crossings, calcifications) on 50 prior/current pairs
CSAW-SScreening mammograms (CC and MLO) of 172 patients; between subjects and between views172 patients (images per patient unverified)Semantic segmentation of cancer and anatomy (pectoral muscle, nipple, mammary gland, lymph nodes...) by experts and non-expertsboth imagesrequestCC BY-NC-ND 4.0, academic and educational use only (access request conditions)non-commercialneeds a request
the data needs a request or a data use agreement
Anatomy masks give inter-subject overlap checks (pectoral muscle, nipple). Licence checked on 2026-10-10 at https://zenodo.org/records/4030660; derived test pairs shareable: unclear; commercial use: no. Files restricted on Zenodo, granted on request; ND bars sharing adapted copies outside the team.
CTI tri-modal ocular tumour set: FA, ICGA and ocular ultrasound (Beijing Tongren)fluorescein angiography, indocyanine green angiography and B-scan ultrasound of choroidal melanoma, haemangioma and metastasis, same patient750 patients; about 518 FA, 486 ICGA, 547 US studies (not all patients have all)none for registration (diagnostic labels)directCC BY 4.0opentruth to make
FA and ICGA of one eye without correspondences (CC BY 4.0); bifurcations found by program, a person checks
FA-ICGA pairs from one session are near-aligned (simultaneous SLO capture likely); ultrasound is not registrable to angiographyvessel bifurcations on FA and ICGA by a program, person verifies; simultaneous FA/ICGA frames, if confirmed, are aligned by construction
CXR-LT (long-tailed CXR classification)MIMIC-CXR frontal/lateral images relabelled for long-tailed multi-label classification (2023, 2024 editions)Same images as MIMIC-CXR (about 377,000, unverified)Image-level labels for 26-45 findings; nothing spatialone imagecontrolledPhysioNet credentialed (unverified)askneeds a request
the data needs a request or a data use agreement
Fails: adds no pairs or spatial truth beyond MIMIC-CXR; listed because it was a lead. Host not re-checked (search budget used up).Use MIMIC-CXR pairs instead (see MIMIC-CXR entry)
Damaged_Section ssEM (Expected Affine)2D serial-section SEM pairs where one section has a fold or crack (mouse SCN, zebrafish, fly optic lobe); neighbouring sections5 pairsfold/crack masks only; no correspondencesone imagedirectnone (no licence file in the repo)unknownleft out
5 downscaled section pairs, no correspondences, no licence
Tiny and downscaled; fails size and truth criteria but is a rare damaged-section test. Licence checked on 2026-10-10 at https://github.com/TongXin-CASIA/Damaged_Section; derived test pairs shareable: unclear; commercial use: unclear. GitHub reports no licence; README has no terms.Two people place 20 landmarks per pair on both sides of the fold; or add synthetic folds to clean sections with a known field
Dark-field chest radiographs, inspiration and expiration (TUM COPD study)Grating-based dark-field and attenuation chest radiographs, PA and lateral, at full inspiration and expiration; same patient87 image pairs (95 participants)Manual lung masks and 6 landmarks by one radiologistboth imageshiddenNot statedunknownleft out
its images or its truth are not released
Private prototype data; CT emphysema scores availableAttenuation and dark-field images come from the same exposure, aligned by construction: a known warp gives exact multimodal truth
Deep compressed multichannel AOSLO (DCAOSLO) data, Duke12-channel non-confocal and confocal AOSLO of healthy and diseased eyes; offset-aperture vs DCAOSLO back-to-back on the same regionunverified countnoneone imagedirectunverified (Duke Research Data Repository)unknowntruth to make
no correspondences; cone centres matched between sessions by program, a person checks; licence unverified
channels from one acquisition are aligned; OA vs DCAOSLO sessions need registrationcone centres matched between OA and DCAOSLO images, person verifies
DeepDRiD (dual-view regular fundus)Colour fundus, disc-centred and fovea-centred views of the same eye; plus separate UWF set~1,000 dual-view pairs (2,000 images, 500 patients); 256 UWF imagesDR grades and quality labels onlyone imagedirectCC BY-SA 4.0opentruth to make
disc- and fovea-centred views of one eye (CC BY-SA 4.0), like FIRE's P pairs, no points: two people place 10 bifurcation pairs on 30 pairs
ISBI 2020 challenge. Partial-overlap pairs like FIRE class P. Fails truth.Vessel bifurcations found by a program and verified by a person in the overlap of 30 pairs.
DenPAR periapical radiographsIntra-oral periapical radiographs, Dental Hospital Peradeniya (Sri Lanka); between subjects1,000 imagesTooth masks with FDI numbers, CEJ and apex keypoints, alveolar crest linesboth imagesdirectCC BY 4.0opentruth to make
no repeat image of a tooth; inter-subject only within one tooth region, or a known projective warp for a subtraction-radiography task
Inter-subject only for same tooth region; no repeat images of a tooth.Subtraction-radiography pairs do not exist publicly: simulate by known projective warp (beam angle change) of one image
Dental caries in bitewing radiographs (Kybic)Bitewing radiographs; between subjects100 imagesCaries boxes by 8 dentists (5 experienced)one imagedirectCC BY-NC 3.0non-commercialleft out
single bitewings with caries boxes: no repeat image, no anatomical correspondence
Fails: no anatomical correspondence, no repeat images. Licence checked on 2026-10-10 at https://data.mendeley.com/datasets/4fbdxs7s7w/1; derived test pairs shareable: yes; commercial use: no. Mendeley Data licence field.Known projective warp of each image (simulated beam angle change) for a synthetic subtraction-radiography task
DENTEX (MICCAI 2023)Panoramic dental radiographs from 3 Turkish hospitals (VistaPano S), patients aged 12+; between subjects693 quadrant, 634 enumeration, 1,005 fully labelled images + 1,571 unlabelledQuadrant and FDI tooth boxes, abnormal-tooth diagnosis boxes (caries, deep caries, periapical lesion, impacted)both imagesdirectCC BY 4.0 (Zenodo)opentruth to make
tooth boxes only; tooth masks from a public model, checked by a dentist on the test images
Boxes only; tooth box centres with FDI numbers give coarse inter-subject points.Tooth instance masks from an OdontoAI-trained model, checked by a dentist on 50 images
Derm7pt (seven-point checklist dataset)clinical and dermoscopic image of the same lesion1,011 cases (>2,000 images)none spatial (7-point criteria and diagnosis)one imageaccountCC BY-NC-ND 4.0non-commercialneeds an account
the download needs an account
Classic cross-modal pair set; clinical images vary in field of view Licence checked on 2026-10-10 at https://derm.cs.sfu.ca/Download.html; derived test pairs shareable: unclear; commercial use: no. Login from a password request form; ND bars sharing adapted copies (derived pairs) outside the team.two people click 4-6 landmarks on 30 pairs
Diabetic colour fundus to FA keypointscolour fundus to fluorescein angiography, retina596 to 8 expert keypoints per pairboth imagesaccountkeypoints CC BY 4.0 (IEEE DataPort); the images come from the Isfahan MISP site, which states no licenceaskneeds an account
the download needs an account
status change: URL found (DOI 10.21227/4hrd-p580, updated 2024-11-19); two annotators confirmed; source images free at misp.mui.ac.ir/en/node/1498. status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Retinal registration review 2024). Used in 3 of the 2D and microscopy registration papers read (KPVSA-Net 2022 (14 pairs); Cross-modal large-FoV 2025 (59 pairs); Modality-invariant coarse-to-fine 2026 (29 pairs)). Licence checked on 2026-10-10 at https://api.datacite.org/dois/10.21227/4hrd-p580; derived test pairs shareable: unclear; commercial use: unclear. Ground truth is open, but pairs need the MISP images (misp.mui.ac.ir/en/node/1498) whose terms are unstated.
Digital Hand Atlas + 37-landmark annotations (Payer)Left-hand radiographs, children; inter-subject895 annotated images37 landmarks on fingertips and joints (annotated by Payer et al.)both imagesdirectlandmark annotations GPL-3.0 (repository licence); images (c) 2008 USC Image Processing and Informatics Lab, no licenceaskin the corpuspriority 2
In the corpus: 724 children of Payer's 895 annotated images (one image each, the repeat images of children followed for years left out), 362 pairs, 30 test and 30 development, 37 landmarks a hand.
The IPILab page (ipilab.usc.edu) links the whole atlas on Google Drive under 'For Research and Education Purpose Only', no licence; landmarks GPL-3.0 in Payer's repository; derived test pairs shareable: unclear; commercial use: unclear.
Digital Mammography DREAM Challenge (Kaiser Permanente Washington)screening mammograms with previous negative exams of the same women (sub-challenge 2)640,905 images, 146,371 exams of 86,873 women; 500-image pilot set publiccancer outcome within one year; no lesion locationsone imagehiddenchallenge data was only usable inside the challenge cloud; pilot data on Synapseaskleft out
its images or its truth are not released
fails access; the 500-image pilot is the only public partnone practical without access
DMR-IR breast thermography databaseBreast infrared thermograms: dynamic protocol of 20 frontal frames over 5 min (motion between frames) plus static views at 0, 45, 90 degrees; same patientabout 290 patients (count unverified); 20-frame sequences eachNone for geometry; diagnosis labels and clinical dataaccountsite conditions of use: research purposes only, cite the group's publications; account requirednon-commercialneeds an account
the download needs an account
Fails truth criterion. Known data anomalies (duplicated frames) documented in Health Informatics J 2023. Licence checked on 2026-10-10 at https://visual.ic.uff.br/dmi/register.php; derived test pairs shareable: unclear; commercial use: no. No named licence; HF mirror SemilleroCV/DMR-IR states none. Research-only clause rules out a paid service.Two people click 10 landmarks (nipples, inframammary fold ends, sternal notch) on first and last frames of 30 sequences.
DRIVE / STARE / HRF / CHASE_DB1 with synthetic warpsSingle colour fundus images with manual vessel masks (and DRIVE bifurcation/crossing labels), used to train retinal keypoint detectors133 images (no real pairs)Manual vessel masks; junctions derived from masks (SuperJunction) or labelled (ConKeD, DRIVE)both imagesdirectvaries per source: HRF CC BY 4.0; DRIVE page states no licence; STARE and CHASE_DB1 hosts unreachable at checkaskleft out
single images: a warp of one image against itself, not a pair; licences vary
Not registration datasets: listed because registration papers use them with synthetic homographies. Used in 2 of the 2D and microscopy registration papers read (ConKeD 2024 (DRIVE, training); KPVSA-Net 2022 (HRF, training)). Licence checked on 2026-10-10 at https://www5.cs.fau.de/research/data/fundus-images/; derived test pairs shareable: unclear; commercial use: unclear. Composite row; only HRF has a confirmed open licence. Decide per source if used.Known smooth deformation plus intensity/illumination change applied to each image; warped vessel masks and junctions give exact truth.
DRTiD (DR two-field image dataset)Colour fundus, macula-centred and disc-centred field of the same eye, same visit (overlap/mosaic registration)1,550 two-field pairs (3,100 images)DR/DME grades and optic disc and macula positions by 3 ophthalmologists; no correspondencesboth imagesrequestopen access for research only; application form, cite Hou et al. 2022non-commercialneeds a request
the data needs a request or a data use agreement
Shanghai Diabetic Eye Study; disc/macula points give a 2-point check only. Fails point truth. Licence checked on 2026-10-10 at https://github.com/FDU-VTS/DRTiD; derived test pairs shareable: unclear; commercial use: no. No licence file; README says research only, which rules out a paid service.Two people place 10 bifurcation pairs in the overlap of 30 pairs; disc/fovea coordinates as extra check.
DSCA and DIASdigital subtraction angiography sequencesmanyvessel masks onlyone imagedirectCC BY 4.0 (both Zenodo records: DSCA 10.5281/zenodo.11255024, DIAS 10.5281/zenodo.11637181)opentruth to make
DSA runs with vessel masks only; DIAS is open on Zenodo, DSCA is restricted; two people click 10 bifurcations on mask and contrast frames of 10 runs
Licence checked on 2026-10-10 at https://api.datacite.org/dois/10.5281/zenodo.11255024; derived test pairs shareable: yes; commercial use: yes. Rights read from DataCite for both records (Zenodo API was rate-limiting); DIAS code repo is Apache-2.0. File access level on Zenodo not rechecked.Two people click 10 vessel bifurcations on mask and contrast frames of 10 DSA runs; vessel masks on both frames as a weaker check.
Duke AOSLO cone detection sets (confocal and split detector)Adaptive-optics SLO confocal and split-detector images of cone mosaicsvalidation and training image sets (counts unverified)Manual cone coordinates by gradersboth imagesdirectGPL v2 (repo); Duke page: non-commercial research onlynon-commercialleft out
detection tiles under non-commercial terms; the Penn AO sets (CC BY) cover AO pairs
Detection data, not registration pairs.Known smooth warp applied to a real tile, cone coordinates warped as exact point truth.
Duke CCM frame registration set (ISATR-Net)Corneal confocal microscopy frames with large misalignment, same eyeunverifiedKeypoint pairs per frame pair (loader expects keypoints_fixed/moving)both imageshiddencode only; data not releasedunknownleft out
its images or its truth are not released
Ask Farsiu lab (Duke) for the data.
Duke/Isfahan FA video and late FA in DME (Rabbani 2015)Fluorescein angiography frame series (early to late) and late images, DME, same eye24 FA videos (768x768) plus late imagesManual and automatic leakage markings on late images; no correspondencesone imagedirectDuke VIP Lab terms: research and educational use; commercialization and redistribution of the images prohibited; cite Rabbani et al. IOVS 2015non-commercialtruth to make
FA frame series with leakage marks only; bifurcations tracked by program, a person checks; redistribution banned (local build only)
Frame-to-frame FA registration with dye-phase intensity change. Fails truth. Duke URL not opened. Licence checked on 2026-10-10 at https://people.duke.edu/~sf59/Rabbani_IOVS_2014_dataset.htm; derived test pairs shareable: no; commercial use: no. (C) 2015 Duke University; redistribution ban covers any sharing of image copies.Vessel bifurcations tracked by a program across frames and verified by a person on 5 frames per video.
e-ophtha (TeleOphta)Colour fundus, screening; some eyes with 2 exams~119 eyes with 2 examsLesion masks (microaneurysms, exudates), no correspondencesone imagerequestaccess link by request form; copyright terms in a Copyright.txt inside the download (not public)askneeds a request
the data needs a request or a data use agreement
Fails truth. Licence checked on 2026-10-10 at https://www.adcis.net/en/third-party/e-ophtha/; derived test pairs shareable: unclear; commercial use: unclear. Page says the database is designed for scientific research and asks for citation; actual terms only visible after the form.Two people place 10 vessel bifurcations on both exams of 20 eyes; AutoMorph vessel junctions verified by a person speed it up.
EchoNet-Dynamic2D apical 4-chamber echo videos, ED to ES10,030expert LV tracings at ED and ESboth imagesaccountStanford research use agreement, non-commercialnon-commercialneeds an account
the download needs an account
status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Foundation models review 2026). Used in 1 of the 2D and microscopy registration papers read (EchoDiST 2026).
EMBED (Emory Breast Imaging Dataset) and EMBED Open DataScreening and diagnostic mammograms (2D, C-view, DBT), Emory 2013-2020, women with yearly priors3.4M images from about 110,000 patients; open release is 20% of 2D and C-view60,000 annotated lesions linked to pathology outcomes (ROI form unverified); priors available by patientone imageaccountEMBED Research Use Agreement: academic non-commercial research only; no distribution, no derivative works, every user registers individuallynon-commercialneeds an account
the download needs an account
Base data of MammoTracker (already listed). ROI on a current lesion and its prior is not guaranteed. status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (MGRegBench 2025). Licence checked on 2026-10-10 at https://github.com/Emory-HITI/EMBED_Open_Data/blob/main/EMBED_license.md; derived test pairs shareable: no; commercial use: no. Clause 3/6: colleagues must register themselves; clause 7 bars derivative works, so derived pairs are not allowed without written permission.Pick 50 prior/current pairs; two readers mark nipple, pectoral muscle line points and calcification clusters visible on both
Emory Knee Radiograph (MRKR)Knee radiographs 2002-2021 from 4 Emory hospitals, many patients with repeat exams (pre/post-op included)503,261 radiographs, 169,578 exams, 83,011 patientsMetadata (laterality, view, hardware), clinical codes, model-inferred KL; no landmarksone imageaccountCC BY-SA 4.0 (GitHub README and AWS Open Data registry)openneeds an account
the download needs an account
Largest longitudinal knee source; 40% African American. Licence checked on 2026-10-10 at https://github.com/Emory-HITI/MRKR; derived test pairs shareable: yes; commercial use: yes. Download goes through registration at data.hitilab.com, which may add its own agreement (not checked, needs an account); ShareAlike on derived pairs.Pick 50 patients with repeat AP views; KNEEL landmarks verified by two people
FAU IR, en-face OCT and OCTA retinal triplets (KPVSA-Net 2022)Infrared fundus, en-face OCT and en-face OCTA (SVP) of the same eye; multimodal, same subject30 test pairs per modality pair (IR-OCT, IR-OCTA, OCT-OCTA)6 manual control points per test image by the authorsboth imageshiddenNot statedunknownleft out
its images or its truth are not released
Private, Department of Ophthalmology, FAU Erlangen; same paper uses public CF-FA (Hajeb) and HRFOCT and OCTA en-face come from the same scan, so they are aligned by construction; a known warp gives exact truth
FETAL_PLANES_DB2D fetal ultrasound standard planes (brain, abdomen, femur, thorax, cervix), same patient repeated images and some longitudinal visits; between patients for atlas-style registration12,400 images from 1,792 patients (2,087 studies); several images of the same plane per patientplane class labels by one expert clinician onlyone imagedirectCC BY 4.0opentruth to make
same-plane fetal ultrasound images of one patient with class labels only (CC BY 4.0); two people place 6-8 plane landmarks on 50 pairs
Fails truth: no landmarks or masks; patient IDs let you pair images of the same planetwo people place 6-8 plane landmarks (e.g. cavum septi pellucidi, thalami, midline ends) on 50 same-plane pairs; or known smooth warp of real images
FI-LOREfundus pairs83landmarksboth imagesavailability unverifiedunknownneeds a request
83 longitudinal fundus pairs, 78 eyes, 54 patients, 10 FIRE-style points a pair, but 'made available by the authors' only (Zhejiang, Juan Ye): ask
FIMD fundus myopia developmentcolour fundus, same eye, years apart during myopia development70 pairs12 control points per pair at vessel bifurcations, with ophthalmologists (FIRE protocol)both imagesaccountCC BY 4.0openneeds an account
the download needs an account
Longitudinal, no other retinopathy.
FIREFLY (fundus registered with FA, Kookmin)Colour fundus vs fluorescein angiography sequences, same eyeunverifiedFA registered to fundus by the authors; fine vessel masks and A/V labels validated by cliniciansone imagerequestno licence; only FIREFLY-Gen has a request form, for academic researchersaskneeds a request
the data needs a request or a data use agreement
Access to FIREFLY itself unclear (form shown for FIREFLY-Gen). Licence checked on 2026-10-10 at https://ailab.kookmin.ac.kr/research/firefly-dataset/request/; derived test pairs shareable: unclear; commercial use: unclear. Access route for the real FIREFLY pairs is not given; the synthetic FIREFLY-Gen is academic-only on request.Two people place 8 bifurcation pairs on 20 pairs; vessel masks as weak check.
FIVES (Fundus Image Vessel Segmentation)Colour fundus photographs, multi-disease, single images (no pairs)None: 800 single imagesPixel-wise vessel masks by crowdsourced medical expertsone imagedirectCC BY 4.0openin the corpus
In the corpus as a segmentation dataset (FIVES); no image pairs for registration.
Fails the pairs criterion; only useful to make synthetic warps or train vessel-based registrationApply known smooth deformations to each image and its vessel mask, exact truth for mono-modal fundus
FLoRI21ultra-widefield fluorescein angiography over time, retina15expert control pointsboth imagesaccountCC BY 4.0openneeds an account
the download needs an account
status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Retinal registration review 2024). Used in 2 of the 2D and microscopy registration papers read (RetinaRegNet 2024; PDM 2026). Licence checked on 2026-10-10 at https://api.datacite.org/dois/10.21227/ydp8-zf19; derived test pairs shareable: yes; commercial use: yes. IEEE DataPort open-access record; download needs a free IEEE account.
FOCTAIRFA vs OCTA, retinal vein occlusion; same eye172 pairs, 29 eyesRegistration ground truth (type unverified)requestno primary page found (VARPA research pages now 404; FOCTAIR not on the archived ophthalmology page)unknownneeds a request
the data needs a request or a data use agreement
Ask VARPA. Licence checked on 2026-10-10 at http://www.varpa.es/research/ophtalmology.html; derived test pairs shareable: unclear; commercial use: unclear. Earlier screening noted 'on demand'; ask the VARPA group (A Coruna) directly.Ask VARPA what the truth is; if none, two people mark 10 vessel bifurcations on FA and OCTA of 20 pairs.
FOVEA (preoperative and intraoperative fundus)preoperative colour fundus photo and intraoperative biomicroscopy frame of the same eye40 patientsvessel and optic disc masks by two clinical fellows on both imagesboth imagesdirectunverified (figshare)unknowntruth to make
vessel and disc masks by two clinical fellows on both images (CC BY, 3.1 GB), but vessel Dice is 0.52-0.68 between the fellows and the views differ 2.4x in scale: make bifurcation pairs from both masks, a person checks
Sci Data 2025. Intraoperative video clips included.vessel bifurcations from both annotators' masks matched by a program then verified by a person
Freiburg longitudinal EOS scoliosis series (Enters 2026)AP whole-spine EOS radiographs of patients 5-25 years with at least 4 follow-ups, some braced, some seated29 patients, 144 exams (88 consecutive pairs)Cobb angles and coronal balance by two radiologistsboth imageshiddenin-houseaskleft out
its images or its truth are not released
brace and posture change between visits: hard deformable spine casevertebral corners on each exam by a detector, verified by a person
FUSeg (Foot Ulcer Segmentation Challenge, AZH)foot ulcer photos from multiple clinical visits, 889 patients1,210 images; visit linkage not releasedwound masks by expertsone imagedirectnone; shared publicly on GitHub with the AZH clinic's permission, no licence fileunknownleft out
visit linkage not released, no licence: no pair to form
Fails pairing: no patient IDs in the release Licence checked on 2026-10-10 at https://github.com/uwm-bigdata/wound-segmentation; derived test pairs shareable: unclear; commercial use: unclear. Repo has no licence; README and challenge README give no terms.known smooth warp of real images with masks
GRAPE (glaucoma longitudinal fundus)Longitudinal colour fundus of glaucoma eyes (Topcon/Canon), same eye across visits631 fundus images from 263 eyes, several visits per eye (pair count unverified)Optic disc and cup contours on every image (expert), no point correspondencesboth imagesdirectCC0opentruth to make
the best retinal set to annotate: 263 glaucoma eyes over several visits, CC0, disc and cup masks only; two people place 10 bifurcation pairs on 20-30 longitudinal pairs
Sci Data 2023. OD/OC labels give only a coarse check near the disc.Vessel bifurcations detected by a program then verified by a person on 30 longitudinal pairs.
H2OPM Image Registration DatasetWWII aerial images vs modern ortho-photo maps, Austria42 pairs (8 references)Manual control pointsboth imagesrequestnon-commercial research use, signed usage agreementnon-commercialneeds a request
the data needs a request or a data use agreement
Non-medical; groupwise Licence checked on 2026-10-10 at https://zenodo.org/records/2533107; derived test pairs shareable: unclear; commercial use: no. Files restricted; access after signing the agreement. Non-medical (historical aerial photos).
HAM10000dermoscopy, several images of the same lesion (lesion_id), often different magnification or angle10,015 images; lesion_id groups repeat imagesnone spatialone imagedirectCC BY-NC 4.0non-commercialtruth to make
repeat dermoscopy of one lesion (lesion_id), non-commercial, no spatial truth: two people click dots and globules on 30 pairs
Repeat images of one lesion can be near duplicates or different zoomtwo people click dots/globules on 30 repeat pairs; known warp of real images
Harvard Whole Brain Atlas fusion pairs (AANLIB CT-MRI, PET-MRI, SPECT-MRI)Pre-aligned 2D brain slice pairs: CT-MR, PET-MR, SPECT-MR; same subject, multimodal810 aligned pairs (184 CT-MR, 269 PET-MR, 357 SPECT-MR) in this mirrorPairs already aligned by the atlas authors; known random warps applied on top give exact truthboth imagesdirect(c) 1995-1999 Johnson and Becker, all rights reserved: personal and educational download and citation allowed; commercial reproduction or multiple distribution prohibited; reproduction needs written permissionnon-commercialleft out
all rights reserved, reproduction needs written permission; 256x256 slices pre-aligned by unknown means
Workhorse of the image-fusion literature; 256x256 slices, few subjects; original alignment quality not documented. status change: Round 4 count: listed in 2 of the 38 round-4 surveys and benchmark papers (Boveiri 2020 review; Han 2022 GAN registration review). Licence checked on 2026-10-10 at http://www.med.harvard.edu/AANLIB/cprt.html; derived test pairs shareable: no; commercial use: no. The GitHub fusion mirror has no licence and itself copies the atlas images.Apply known affine plus smooth B-spline warps to one image of each pair: exact multimodal truth (MatchAnything protocol).
Harvard-FairDomain (paired en-face OCT fundus and SLO fundus)En-face fundus projection from OCT vs SLO fundus image of the same eye (glaucoma clinic); multimodal, same subject10,000 patients in the segmentation subset (one pair each); 10,000 more in the classification subsetCup and disc masks on both images: from OCT device software, carried to SLO by NiftyReg and checked by experts (80% accepted)both imagesdirectCC BY-NC-ND 4.0 (repo); also on HF harvardairobotics/FairDomain and Zenodo 13178701non-commercialtruth to make
its SLO masks were carried by NiftyReg (circular) and cover the disc only; two people click 10 bifurcations on 50 pairs; NC-ND
SLO masks were made by registration, so they are circular as registration truth; cup/disc only covers the optic nerve head. ND clause may block redistributing derived pairs.Two people click 10 vessel bifurcations on both images for 50 pairs; use the masks only as a coarse Dice check.
Harvard-GDP (glaucoma detection and progression)OCT RNFL thickness maps (225x225), one per patient, with visual fields and progression labels; between subjects only1,000 maps (1,000 patients), no repeat visits releasedGlaucoma and VF progression labels onlyone imagedirectCC BY-NC-ND 4.0non-commercialleft out
one thickness map per patient, no repeat visit
Hosted on Hugging Face (harvardairobotics/Harvard-GDP). Fails pair and truth criteria.Vessel shadows on RNFLT maps as weak landmarks for between-subject alignment; no longitudinal truth possible.
HIBA Skin Lesions (Hospital Italiano de Buenos Aires)clinical and dermoscopic images of the same lesions, Argentina~1,600 images; lesion_id links clinical and dermoscopic viewsnone spatialone imagedirectCC BYopentruth to make
clinical and dermoscopic images of one lesion (CC BY), no spatial truth; two people click 4-6 landmarks on 30 pairs
Open-licence clinical-dermoscopic pairs (smaller than MILK10k); two DOIs (2019-2022 release and earlier)two people click 4-6 landmarks on 30 pairs
High-resolution fundus pairs for thin-feature registration (Zenodo 4304781)grayscale fundus image pairs built from HRF images by elastic deformation, same eye42 pairsknown elastic deformation (whether the field is shipped is unverified) and expert vessel labelsboth imagesdirectCC BY 4.0openleft out
pairs made by a published elastic warp of HRF images: a method can learn the generator; the corpus draws its own unpublished warps
synthetic warps of real images; vessel labels come from HRF expertsif fields are missing, regenerate the warps; vessel bifurcations give point truth
Hip FAI landmarks, paired X-ray and MRI (Oxford FAIT)AP pelvis radiograph and coronal hip MRI slice of the same patient (FAIT trial); same subject, two modalities89 patients, 162 image pairs4 landmarks per hip (femoral head centre, neck axis point, lateral acetabular edge, cam point) on both X-ray and MRIboth imagesrequestdata not released; repository code MITunknownneeds a request
the data needs a request or a data use agreement
Projection vs slice, so only in-plane landmarks match; data availability unverified. Licence checked on 2026-10-10 at https://github.com/Malga-Vision/Landmarks-Hip-Conditions; derived test pairs shareable: unclear; commercial use: unclear. README describes the FAIT-trial cohort (Oxford, 89 patients) but gives no access route or terms for the images.Add 6 more matched landmarks per hip; use only the femoral head centre and acetabular edge as cross-modal check
HKU Duchess of Kent AIS curve progression cohort (Wang 2024)standing PA spine and hand X-rays of adolescents with idiopathic scoliosis followed every 3-6 months to skeletal maturity710 patients with follow-up seriesCobb angle trajectories, maturity grades; released images are fixed-size ROI cropsone imagehiddenCC BY-NC 4.0 on the record, but no files listed (restricted or embargoed)non-commercialleft out
its images or its truth are not released
if released, only JPG ROI crops around the apex and the global spine: geometry partly lostvertebral endplate corners by a detector plus person on first and last visit
Hong Kong EOS AIS cohort with endplate landmarks and follow-ups (QMH/DKCH, 2025)EOS biplanar whole-spine PA radiographs of adolescents with idiopathic scoliosis, two centres630 patients; 138 with follow-up exams4 endplate corners on 18 vertebrae (C7-L5) by two orthopaedic surgeonsboth imageshiddenin-houseaskleft out
its images or its truth are not released
if shared, follow-up pairs with 72 landmarks each would be a strong 2D spine registration set
HRF image-quality pairs (FAU Erlangen)Colour fundus (Canon CR-1), same eye captured twice: blurred then repeated good image18 pairsQuality labels only; small eye-movement shift between capturesone imagedirectpage releases the HRF database under Creative Commons 4.0 Attribution, free for research; the image-quality section itself only asks for a citationasktruth to make
18 eyes photographed twice (defocused, then good); two people place 10 bifurcation pairs each; small eye motion, so check the room first
Repeat captures with defocus; small motion. Fails truth. Licence checked on 2026-10-10 at https://www5.cs.fau.de/research/data/fundus-images/; derived test pairs shareable: unclear; commercial use: unclear. CC BY sentence sits in the vessel-segmentation paragraph; whether it covers the 18 quality pairs is not explicit. Likely CC BY, worth a one-line confirmation.Two people place 10 vessel bifurcation pairs per pair (18 pairs is quick).
HVAngleEst (foot radiographs, hallux valgus)Weight-bearing dorsoplantar foot radiographs; between subjects, or left-to-mirrored-right foot when both are on one image1,382 images, 1,587 feet, 1,150 patients (204 images show both feet)Two endpoints each for great toe, first and second metatarsal axes, plus foot boxes; four orthopaedic surgeons with adjudicationboth imagesdirectnot visible: the Science Data Bank share link is script-rendered and no DataCite record was found; paper states noneunknowntruth to make
6 axis end points a foot, all on the forefoot, licence not found; add 10 bone landmarks on 30 feet
Only 6 axis endpoints per foot, all on the forefoot; no follow-up pairs. 50 images are phone photos of films. Licence checked on 2026-10-10 at https://www.scidb.cn/en/s/FVFFnq; derived test pairs shareable: unclear; commercial use: unclear. Many Science Data Bank records are CC BY 4.0 (e.g. CSXA), but this one could not be confirmed.Add 10 bone landmarks (navicular, cuboid, calcaneal tuberosity, metatarsal heads) by two people on 30 feet for inter-subject TRE.
HVDROPDB (ROP RetCam/Neo, multi-view per eye)Infant fundus (RetCam, Neo), 2-12 views per eye (posterior, temporal, superior, inferior, nasal) for mosaickingmulti-view sets per eye (count unverified); ~550 MBOptic disc, vessel and ridge masks by ROP experts on a subset; no correspondencesone imagedirectCC BY 4.0opentruth to make
multi-view infant fundus without correspondences; 8 bifurcation pairs in the overlap of 20 view pairs (COph100 already covers infants)
COph100 authors rejected it (poses only, no time series). Whether views of one eye are grouped in files is unverified.Vessel masks on both views plus 8 manual bifurcation pairs in overlaps of 20 view pairs.
IAC-TecMed thermography database (IAC-TMBD)Sole of the feet: thermal infrared + RGB + depth at T0, T5, T10, T15 min; multimodal (IR to RGB-D) and temporal, same subject19 healthy volunteers x 4 time points (RGB-D-IR), raw unregistered RGB-D includedFoot masks drawn by three researchers merged with STAPLE, on IR-registered images; raw RGB-D unregisteredboth imagesrequestCC BY 4.0openneeds a request
the data needs a request or a data use agreement
Masks give Dice-only truth; the rig calibration gives a near-known IR-RGB transform. Licence checked on 2026-10-10 at https://www.iac.es/en/science-and-technology/technology-transfer-iactec/forms; derived test pairs shareable: yes; commercial use: yes. Download after giving e-mail and affiliation and accepting a notice; licence line on the form page: 'IAC-TecMed Database by Instituto de Astrofisica de Canarias is licensed under CC BY 4.0'.Use the provided processed (registered) set as reference and the raw set as moving: known transform per frame; add 8 clicked contour landmarks per foot.
IARC Colposcopy Image Bankcolposcopy photos of the same cervix after saline, after acetic acid, after Lugol's iodine and with green filter (contrast change between stages)200 cases, 913 images (new VIA/colposcopy bank: 5,000 cases, 56,000 images)none spatial (Swede score, diagnosis metadata)one imagerequestfree to researchers via access request form; IARC/WHO non-exclusive licence for selected published materials; no sharing of the data without further permissionaskneeds a request
the data needs a request or a data use agreement
Real multi-stain-like pairs: acetowhitening and iodine change appearance strongly; camera moves between stages Licence checked on 2026-10-10 at https://screening.iarc.fr/cervicalimagebank.php; derived test pairs shareable: no; commercial use: unclear. Page names 'regulated research groups or commercial developers of AI' as intended users, so commercial use may be possible; the licence text itself is not linked.two people click 8-10 landmarks (os, vessels, cysts, transformation zone border) on 50 case pairs
IICS-UNA bone scan dataset (Paraguay)Tc-99m bone scintigraphy anterior and posterior whole-body views, breast cancer patients; ANT-POST within patient or between patients291 patients, 582 imagesDiagnosis label onlydirectCC BY 4.0 (Zenodo API)opentruth to make
anterior and posterior bone scans with a diagnosis only (CC BY 4.0); two people click 15 skeletal landmarks on 20 patients
Fails truth criterion; small open add-on to BS-80K.Skeleton masks from a public bone-scan segmenter on both views (weak), or 15 clicked skeletal landmarks on 20 patients by two people.
INbreastFull-field digital mammograms, Centro Hospitalar de Sao Joao (Porto); CC and MLO of the same breast115 cases, 410 imagesSpecialist lesion contours (masses, calcifications, asymmetries, distortions) in XMLboth imagesrequestno official page or licence found; obtained by contacting the original authorsunknownneeds a request
the data needs a request or a data use agreement
Status change: no official download portal found (2026-10-10); obtained by contacting the authors. status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (MGRegBench 2025). Licence checked on 2026-10-10: no primary page states terms; derived test pairs shareable: unclear; commercial use: unclear. Original portal no longer maintained; terms only from the authors.Lesion contours on both views give CC-MLO correspondence checks
INSPIRE-stereo (Iowa)Stereo optic-disc fundus photo pairs, glaucoma, same eye same session30 stereo pairsDepth reference from same-session SD-OCT registered to the photos (Iowa group)one imagedirectFree for research and education; no redistribution, no commercial usenon-commercialtruth to make
30 stereo disc pairs, OCT depth on one side only; two people mark 10 vessel points on both images; non-commercial, no redistribution
Dense disparity truth via OCT depth; stereo, not temporal.OCT depth covers the photos only one way; two people mark 10 vessel points on both stereo images of the 30 pairs for sparse disparity truth. Non-commercial.
InVivoDataForUSE (forearm RF elastography sequences, KCL)In vivo ultrasound RF sequences of a human forearm under slow probe compression; frame-to-frame displacement, same subject310 sequences, 17,271 frames, one volunteerNone released; the paper used 8 landmark pairs per test sequence (authors)accountnone; Synapse project has no access requirement or licence, wiki asks to cite Delaunay et al. PMB 2021unknownneeds an account
the download needs an account
Single volunteer; test-split landmarks not public. Licence checked on 2026-10-10 at https://www.synapse.org/InVivoDataForUSE; derived test pairs shareable: unclear; commercial use: unclear. Read through the Synapse REST API (project syn25007887): files public to Synapse users, no terms of use attached.Two people click 8 vessel-wall points on first and last frame of 20 sequences.
ISBI 2015 cephalometricLateral cephalograms; inter-subject400 images (400 subjects)19 landmarks by 2 orthodontistsboth imagesdirectCC BY-SA 4.0 (per CL-Detection 2023 paper; original host not checked)openneeds an account
the NTUST host answers 404; the images now come with CL-Detection 2023 on grand-challenge (account and agreement); Aariz covers the problem openly
Classic inter-subject 2D landmark set.
Isfahan MISP FA and SLO pairs (Almasi 2020)Fluorescein angiography (30 and 55 deg) vs infrared SLO, Heidelberg Spectralis HRA2/OCT, DR, same eye36 pairs, 21 subjectsNone on page (COph100 review lists registration GT, private; unverified)directnone stated on the MISP dataset pageunknowntruth to make
36 FA-SLO pairs of 21 subjects with no released truth (the authors' points are private); two people place 8 bifurcation pairs; no licence stated
SLO is aligned by construction to OCT on Spectralis, so FA-SLO doubles as FA-OCT. Licence checked on 2026-10-10 at https://misp.mui.ac.ir/fa/fa-and-slo-images; derived test pairs shareable: unclear; commercial use: unclear. Isfahan MISP research centre page gives the download links and the source paper, no licence or terms of use.Two people place 8 bifurcation pairs per pair; ask authors for their evaluation points.
Isfahan MISP fundus to OCT projection registration (22 pairs)Colour fundus vs OCT en-face projection (Topcon 3D OCT-1000), retinal disease, same eye22 pairs (17 macular, 5 peripapillary)None stated on pagedirectnone stated on the MISP dataset pageunknowntruth to make
22 fundus/OCT-projection pairs with no truth; two people place 6-10 bifurcation pairs; projections computed from the volumes; no licence stated
Two OCT links mismatched on page. Projection must be computed from OCT volume. Licence checked on 2026-10-10 at https://misp.mui.ac.ir/fa/database-22-retinal-images-purpose-vessel-based-registration-fundus-and-oct-projection-images; derived test pairs shareable: unclear; commercial use: unclear. Isfahan MISP research centre page gives the download links and the source paper, no licence or terms of use.Two people place 6-10 vessel bifurcation pairs per pair.
Isfahan MISP Heidelberg corneal OCT (15 subjects)Anterior-segment (corneal) OCT B-scan stacks, Heidelberg; neighbouring B-scans of the same cornea15 subjects (e.g. 41 B-scans of 240x748 for subject 1)None stated (paper segments corneal layer boundaries)directnone stated on the MISP dataset pageunknownleft out
neighbouring B-scans of one volume: slices, not a registration pair
Not repeat captures; slice-to-slice only. Fails pair and truth criteria. Licence checked on 2026-10-10 at https://misp.mui.ac.ir/fa/node/1397; derived test pairs shareable: unclear; commercial use: unclear. Isfahan MISP research centre page gives the download links and the source paper, no licence or terms of use.Two people trace epithelium/Bowman/endothelium boundaries on 5 B-scans per subject; known warp on a B-scan for exact truth.
Isfahan MISP OCT and colour fundus of 50 healthy eyesColour fundus vs OCT volume (projection), left and right eyes of 50 healthy persons~100 eyesNoneone imagedirectnone stated on the MISP dataset pageunknowntruth to make
fundus and OCT volumes of 50 people without truth; bifurcation pairs on 20 eyes; no licence stated
Unpaired truth; a CF to OCT-projection set. Fails truth. Licence checked on 2026-10-10 at https://misp.mui.ac.ir/fa/node/1388; derived test pairs shareable: unclear; commercial use: unclear. Isfahan MISP research centre page gives the download links and the source paper, no licence or terms of use.Two people place 6 bifurcation pairs on 20 eyes.
ISIC reflectance confocal microscopy lesions (dermoscopy + RCM)dermoscopic image vs RCM macroscopic camera image vs RCM mosaics and tiles of the same lesion (cross-device, cross-scale; tile-to-mosaic)5 lesions (48 RCM mosaics; 8-15 images per lesion incl. dermoscopy)none spatial; RCM tiles composing mosaics give stitching structure (unverified)one imagedirectCC BYopenleft out
5 lesions: fewer than ten independent pairs
Tiny but a real dermoscopy-to-confocal registration case used clinically to locate RCM mosaicstwo people click 6 landmarks between dermoscopy and RCM macroscopic image on all 5 lesions
iToBoS 2024 skin lesion detection (3D-TBP tiles)skin-region images cut from 3D total-body photographs, lesion detection16,954 tiles; overlapping tiles of the same patient unverifiedlesion bounding boxesdirectCC BYopenleft out
detection tiles whose overlaps are not documented
Pairing between tiles not documented; mainly a detection setif overlapping tiles exist, box centres of the same lesion as correspondences; otherwise known warp of a real tile
iToBoS skin region images from 3D total body photography2D skin-region crops rendered from 3D total body photographs (Canfield Vectra), 100 participants16,954 images; overlapping regions of the same participant (pair count unverified)Lesion bounding boxes verified by dermatologists; no correspondence across imagesone imageaccountCC BY-NC-SA 4.0 (paper); Kaggle competition rulesnon-commercialneeds an account
the download needs an account
Fails truth for registration; 3D meshes not releasedPair overlapping crops of the same participant and match the boxed lesions (two annotators) for 2D pair truth
Jordan (KAUH) scoliosis, spondylolisthesis and normal spine X-rays (Fraiwan 2022)spine radiographs of 338 subjects (188 scoliosis, 79 spondylolisthesis, 71 normal), one exam eachno follow-ups; between-subject onlydiagnosis labels onlydirectCC BY 4.0 (Mendeley default, unverified)openleft out
single exams, diagnosis only; CSXA and BUU-LSPINE cover spine radiographs with expert points
fails pairs and truth; listed as the only public scoliosis set found besides AASCEvertebra corner landmarks from a public detector, checked by a person, for between-subject tests
JSRT chest radiographsPA chest radiographs; inter-subject247 imagesNodule positions only (organ masks come from SCR)requestfree of charge but not public domain: do not distribute any images, no commercial usenon-commercialneeds a request
the data needs a request or a data use agreement
Base images for SCR. status change: Round 4 count: listed in 6 of the 38 round-4 surveys and benchmark papers (Boveiri 2020 review; Bharati 2022 review; Hammoudeh 2023 survey; Chen X 2021 review (Prog Biomed Eng); Azad 2023 ViT review; Han 2022 GAN registration review). Licence checked on 2026-10-10 at http://db.jsrt.or.jp/eng-01.php; derived test pairs shareable: no; commercial use: no. (c) Japanese Society of Radiological Technology; user list kept on registration.Inter-subject only: SCR gives lung, heart and clavicle contours with point landmarks on every image, so pairs get contours and boundary points on both (shape, not exact anatomy).
KBSMC standard to ultra-widefield fundus pairsStandard fundus vs ultra-widefield fundus, same patient, 1-4x scale gap3,744 pairs (374 test)Manual point correspondencesboth imageshiddennot statedunknownleft out
its images or its truth are not released
Kangbuk Samsung Medical Center; no data statement seen. Used in 1 of the 2D and microscopy registration papers read (PDM 2026).
KNEEL OAI knee landmarksBilateral PA knee radiographs from OAI, all KL grades; between subjects748 knee joints16 anatomical landmarks (tibial plateau, femoral condyles) initialised by BoneFinder and refined by a person experienced in knee anatomyboth imagesaccountannotations: no licence file in the repo; images: OAI via the NIMH Data Archive under its data use terms (account)unknownneeds an account
the download needs an account
Annotation file annotations/bf_landmarks_1_0.3.csv is on the master branch only. OAI has 9 years of follow-up radiographs for the same knees. Licence checked on 2026-10-10 at https://github.com/imedslab/KNEEL; derived test pairs shareable: unclear; commercial use: unclear. Landmark CSV has no stated licence; OAI images need an NDA data use certification, not read here.Temporal pairs: take the same 748 knees at 24/48 months in OAI and have two people place the same 16 landmarks
Lee et al. 2015 CF to OCT registration setColour fundus vs OCT fundus (en-face) image, same eye52 pairs8 corresponding points per pairboth imageshiddenno public release statedunknownleft out
its images or its truth are not released
MIFNet describes it without saying it is public; no host found. Ask J. A. Lee / J. Cheng.
LLRFormer full-length lower-limb radiographsFull-length standing lower-limb radiographs, many with metal implants; between subjects1,147 images (709/102/200 + 136 external)36 keypoints (hip, femur, knee, tibia, ankle)both imagesaccount'other' (Hugging Face card), gated: conditions shown only after login and contact-sharing approvalaskneeds an account
the download needs an account
Very tall images (about 8,000 px): good stress test for large 2D registration. Licence checked on 2026-10-10 at https://huggingface.co/datasets/YANG1568279/full_length_lower_limb_radiographs_with_metallic_implants; derived test pairs shareable: unclear; commercial use: unclear. Manual gate; the conditions text is not visible without an account.
Longitudinal healthy multi-modal AO datasets with two timepoints (Chen 2019)AOSLO confocal, split-detector and dark-field montages of the same healthy eyes at two timepoints4 subjects, 2 timepoints, many locationsnone released as point truth (validation set for constellation matching)one imagedirectCC BY 4.0opentruth to make
4 healthy subjects at two timepoints (CC BY 4.0); cones matched by program, a person checks 50 a location; few subjects
cone mosaics are stable in healthy eyes, so individual cones are natural landmarkscone centres detected by a public detector, matched across timepoints, person verifies 50 cones per location
Longitudinal images for lesion viewer testing (ISIC)dermoscopic and clinical images of the same lesions at different days (acquisition_day), matched to longitudinal images already in the ISIC Archive127 images (plus the matched archive images)none spatial (lesion and patient IDs, day offsets)one imagedirectCC BYopentruth to make
longitudinal dermoscopic and clinical images (CC BY) with no spatial truth; two people click 6-10 structures on 30 pairs
The ISIC Archive holds more longitudinal dermoscopy via lesion_id + acquisition_day; this collection is a starting indextwo people click 6-10 structure landmarks per pair on 30 pairs
Longitudinal overview images of posterior trunks (ISIC, UPMC)digital camera photos of the back of melanoma-risk patients at different visits (acquisition_day field)36 images, a few patientsnone spatialone imagedirectCC BYopentruth to make
17 patients photographed at two visits 2-4 years apart (checked in the ISIC API), CC BY; two people click 20-30 matched naevi a pair: a real deformable photo task
Tiny but truly longitudinal total-body-style photos with open licencetwo people click 20-30 matched naevi per pair (the lesions are natural landmarks)
LSFG laser speckle flowgraphy set (RetinaRegNet)Laser speckle flowgraphy retinal blood-flow maps, same eye15 pairsLandmark pairs (as used for AUC/MLE)both imageshiddenprivate; 'to be released soon'unknownleft out
its images or its truth are not released
Private as of repo README. Used in 1 of the 2D and microscopy registration papers read (RetinaRegNet 2024).
Lu-177 DOTATATE multi-time-point planar whole-body scans (Deep Blue)Planar whole-body Lu-177 scintigraphy, anterior and posterior, 4 time points within 8 days, plus SPECT/CT at the same times; same patient over time2 patients x 4 time points x 2 views (3 energy windows)Lesion and organ VOIs on SPECT/CT drawn by a radiologist; nothing on the planar imagesone imagedirectCC BY-NC 4.0non-commercialleft out
2 patients
Planar-to-planar time registration is the dosimetry use case; tiny. Companion records hold the SPECT/CT and segmented VOIs. Licence checked on 2026-10-10 at http://deepblue.lib.umich.edu/data/concern/data_sets/1j92g763g; derived test pairs shareable: yes; commercial use: no. Rights read from DataCite for 10.7302/y4xd-s758.Project the SPECT/CT VOIs to the planar geometry (2D/3D) to get organ masks on every time point; or two readers click 10 organ/lesion centres per scan.
LUTSeg (longitudinal leprosy ulcer tissue segmentation)smartphone photos of the same leprosy ulcer at successive visits over 21 months141 images from 39 patients (3.6 visits per patient)wound masks and five tissue classes per visit; 46-image subset by five cliniciansboth imagesdirectCC BY 4.0 (data)opentruth to make
the same ulcer at successive visits, CC BY 4.0, wound masks change with healing; two people click 6-10 skin landmarks around it on 30 pairs
Wound shape changes with healing, so masks are not a stable registration check; surrounding skin is the stable parttwo people click 6-10 skin landmarks (scars, freckles, hair whorls) around the ulcer on 30 visit pairs
Macular hole surgery longitudinal OCT (figshare 32605218)SD-OCT B-scans before and after macular hole surgery, up to 6 follow-ups over 48 monthsbaseline + 6 postoperative stages; patient count unverifiedpixel-level segmentation (labels and annotators unverified)directCC BY 4.0opentruth to make
foveal H and V B-scans before and up to 48 months after surgery with 8-class masks of unstated origin; check who drew them and whether the scan position repeats
Published 2026-06; 2D B-scans, slice matching across visits not guaranteed.layer boundaries on both as weak check; fovea centre marked on each visit
MARIO (MICCAI 2024)OCT B-scan pairs from consecutive visits (Spectralis), AMD; same eyeBrest part only (Tlemcen private); thousands of slice pairsActivity-evolution labels; pairs already aligned by Spectralis follow-upone imagedirectCC BY 4.0opentruth to make
B-scan pairs already aligned by the Spectralis follow-up; only a known warp gives a truth
Device alignment as weak reference only.Pairs are already aligned by the device: no independent truth; a known shift and small warp of one B-scan gives synthetic truth only.
MATADOR AO flood-illumination montages (Brno)Adaptive-optics flood-illumination retinal images assembled into montages, 10 healthy subjects200 images into 20 montagesPhotoreceptor labels by three graders (subset); montage by automatic pipelineone imagedirectCC BY 4.0 (Zenodo 10.5281/zenodo.6635685, open access)opentruth to make
montages made by an automatic pipeline; two people place 10 point pairs in 20 tile overlaps
Paper says dataset is public; host not confirmed. Licence checked on 2026-10-10 at https://zenodo.org/records/6635685; derived test pairs shareable: yes; commercial use: yes. Host found through the BOE 2023 data availability statement (200 images of 10 subjects plus labelled patches).Two people place 10 point pairs in the overlap of 20 tile pairs.
MeDAL-Retina (keypoint-annotated fundus)Colour fundus images (from e-ophtha, a classification set, FIRE) with manual keypoints, single images208 labelled images (18-86 keypoints each) + 1,920 unlabelledManual keypoints by lab students (single images)one imagedirectCC BY 4.0openleft out
single images with student keypoints (some are FIRE's images): no pair
Not pairs; Dropbox download. Source images keep their own terms.Known warp applied to each labelled image; warped keypoints as exact truth.
MedMNIST v228 to 224 px images and 3D volumes from public setsnoneclass labelsone imagedirectCC BY 4.0 (unverified)openleft out
tiny unpaired images
Fails pairs: one image per subject, no repeat or second modality. Used by ATOM 2024 as a source. Round 4 count: listed in 2 of the 38 round-4 surveys and benchmark papers (Foundation models review 2026; ATOM 2024).Tiny unpaired images: only a known deformation for toy tests.
MEMOEMA to OCTA of the retina (non-human primate)306 expert points per pairboth imagesdirectnone stated (Google Drive download from the project page; cite Wang et al. BOE 2024)unknownleft out
30 EMA-OCTA pairs from four eyes of two macaques, six points a pair: neither ten independent pairs nor ten points
Licence checked on 2026-10-10 at https://chiaoyiwang0424.github.io/MEMO/; derived test pairs shareable: unclear; commercial use: unclear. Project page and repo carry no licence.
MIAS (Mammographic Image Analysis Society) v1.21Digitised film MLO mammograms, left and right breast of the same woman322 images = 161 left/right pairsAbnormality centre and radius; no anatomical landmarksone imagedirectCC BY (Cambridge record); README says research useopentruth to make
161 left/right MLO pairs with abnormality centres only; two readers mark nipple, pectoral line and 5 symmetric points
Contralateral pairs (mirror one breast) are the classic bilateral registration task. status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (MGRegBench 2025).Two readers mark nipple, pectoral muscle line ends and 5 symmetric gland/vessel points on all 161 pairs
MIDRC-RICORD-1CChest radiographs of COVID-19 positive adults, 4 sites, several radiographs per patient998 radiographs of 361 patients (about 2.8 per patient)Exam-level labels and airspace disease grade by thoracic radiologists; no boxesone imagedirectCC BY-NC 4.0non-commercialtruth to make
about 2.8 radiographs a COVID patient, exam labels only, non-commercial; two people place 20 landmarks on 30 pairs
Study dates pseudonymised but order kept (unverified).Two people place 20 landmarks on 30 pairs
MILK10kclinical close-up photo vs dermoscopic image of the same skin lesion (cross-device, cross-scale)5,240 lesion pairs (10,480 images) plus 479-pair benchmarknone spatial (pairing and diagnosis only)one imagedirectCC BY-NC 4.0non-commercialtruth to make
clinical close-up and dermoscopic image of 5,240 lesions, non-commercial; two people click 4-6 landmarks on 50 pairs
Largest clinical-dermoscopic pair set; scale ratio large (close-up vs magnified), lesion often only part of the close-uplesion masks on both (public segmenter, checked) plus two people clicking 4-6 border or colour landmarks on 50 pairs
MIMIC-CXR (longitudinal priors)Chest radiographs, same patient over time (ICU-heavy, BIDMC Boston)377,110 images, 227,835 studies, 65,379 patients (paper); tens of thousands of patients with 2+ studiesNone spatial; free-text reports that compare with the priorone imagecontrolledPhysioNet Credentialed Health Data License 1.5.0askneeds a request
the data needs a request or a data use agreement
Parent of MS-CXR-T, Chest ImaGenome, CheXTemporal, TiBiX MIMIC-T, TRACE and GRCD pair lists; CheXmask gives automatic lung/heart masks on these images. status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Foundation models review 2026).Pick ~50 frontal PA pairs; two people place ~20 landmarks (clavicle heads, carina, rib crossings, costophrenic angles, device tips) on both images; CheXmask masks as a weaker overlap check
Montgomery County CXRPA chest radiographs, TB screening; inter-subject138 imagesManual lung masks (radiologist-supervised)both imagesdirectnone stated; NLM says it made the set publicly available for researchunknowntruth to make
manual lung masks only: after a moment-matching affine the lung Dice between subjects is already 0.89 (measured on 10 pairs), so no room; needs reader landmarks
status change: Round 4 count: listed in 3 of the 38 round-4 surveys and benchmark papers (Hammoudeh 2023 survey; Chen X 2021 review (Prog Biomed Eng); Han 2022 GAN registration review). Licence checked on 2026-10-10 at https://pmc.ncbi.nlm.nih.gov/articles/PMC4256233/; derived test pairs shareable: unclear; commercial use: unclear. Jaeger et al. QIMS 2014 (the NLM release paper) has no licence or terms; Open-i FAQ page now only shows a funding-lapse notice.
MOST knee radiographs (longitudinal)PA fixed-flexion, lateral and full-limb radiographs, baseline to 84 months, same knees over time3,026 subjects with up to 6 follow-ups (Tiulpin 2018)Semi-quantitative readings (KL, JSN, osteophytes, alignment); no landmarksone imagerequestMOST public data sharing with a 'steps to obtaining data and images' process (data use agreement); DUA text not readableaskneeds a request
the data needs a request or a data use agreement
Beam angle varies (5, 10, 15 degrees) between images, a realistic projection change. Licence checked on 2026-10-10 at http://web.archive.org/web/20170603141742/http://most.ucsf.edu:80/data&images.asp; derived test pairs shareable: unclear; commercial use: unclear. most.ucsf.edu did not answer on 2026-10-10; only the archived summary page was read.Run KNEEL/BoneFinder landmarks then verify by two people on 50 knee pairs
MOSTA Stereo-seq mouse embryo atlasStereo-seq sagittal sections of mouse embryos E9.5-E16.5; E16.5 has 13 sections from one embryo; serial-section and between-stage53 sections; 13 serial sections of one E16.5 embryoNone for alignment; annotated tissue clusters per sectionone imagedirectnone stated on the MOSTA pages (cite Chen et al. Cell 2022)unknownleft out
13 serial sections of one embryo and no alignment truth
Whether ssDNA images are included is unverified. Licence checked on 2026-10-10 at https://db.cngb.org/stomics/mosta/download/; derived test pairs shareable: unclear; commercial use: unclear. Download page does include nucleic-acid staining images; STOmicsDB site terms page is script-rendered and was not readable.Tissue cluster labels as weak labels on adjacent sections; known warp of one section for exact truth
MS-CXR-TFrontal chest radiograph pairs (current and nearest prior), same patient, from MIMIC-CXR1,326 image pairs from 800 subjectsProgression labels (improving, stable, worsening) for 5 findings, reviewed by board-certified radiologists; no spatial marksone imagecontrolledPhysioNet Credentialed Health Data License 1.5.0askneeds a request
the data needs a request or a data use agreement
Fails spatial truth; useful as a curated pair list with real interval change.Use the pair list; two people place 20 anatomical landmarks on 50 pairs; change labels say where intensity change is expected (mask those regions from similarity checks)
Multi-centre Ningxia/Shenzhen panoramic datasetPanoramic radiographs from 3 hospitals (impacted teeth, periodontitis, caries); between subjects6,536 images (4,940 + 716 + 880)Disease segmentations and labelsone imagerequestfully shared; authors request no commercial use and citation (no licence file)non-commercialneeds a request
the data needs a request or a data use agreement
Lesion masks only, not anatomical correspondences. Licence checked on 2026-10-10 at https://github.com/qinxin99/qinxini; derived test pairs shareable: unclear; commercial use: no. README line: 'we kindly request that you do not use it for any commercial purposes'.Tooth masks by a public model verified on 30 images
Multi-modal AO datasets with manual montages (Chen 2016)AOSLO confocal, split-detector and dark-field image sets of the same retinal locations; montage of overlapping tiles11 eyes (8 control, 2 CSCR, 1 RP), dozens of tiles eachmanual montage by an expert (Photoshop file)both imagesdirectCC BY 4.0opentruth to make
expert Photoshop montages give tile placements, not a deformation; cone matches by program, a person checks; channels simultaneous (a known warp would be exact)
the three AOSLO channels are acquired simultaneously, so aligned by constructioncone centres on overlapping tiles matched by a program, person verifies; multimodal channels exact by construction
MultiRegEvalfluorescence to quantitative phase; SHG to bright field, cells and tissue5,040 + 536synthetic rigid transformsboth imagesdirectCC BY 4.0opentruth to make
its truth is synthetic rigid moves, which the corpus's alignment removes; the source pairs (SHG and bright field, QPI and fluorescence) are aligned by construction, so a known warp gives a 2D multimodal microscopy truth
rigid only
MURAUpper-extremity radiographs (elbow, finger, forearm, hand, humerus, shoulder, wrist), several views per study40,561 images, 14,863 studies, 12,173 patientsNormal/abnormal per study by radiologists; nothing spatialone imageaccountStanford AIMI research use agreement (not stated on page)non-commercialneeds an account
the download needs an account
Fails: no landmarks; repeat studies per patient exist but views differ.Pick same-view wrist/hand pairs of the same patient; two people place 20 bone landmarks
ND-TimeLapseIris-2012 and ND-Iris-Template-Aging (Notre Dame)Near-infrared iris images of the same eyes over months to yearsTimeLapse: 6,797 images, 46 irises (2004-2008); Template-Aging: 11,776 images (2008-2010)Identity labels onlyone imagerequestUND Biometrics Database Release Agreement signed by the institution: internal research only; no redistribution (even to another unit of the same organisation), no modification or commercial use without approvalnon-commercialneeds a request
the data needs a request or a data use agreement
Fails truth; pupil dilation changes make non-rigid iris warps. Licence checked on 2026-10-10 at https://cvrl.nd.edu/media/django-summernote/2018-09-19/880510cc-5001-4025-b2af-79f8e43324b9.pdf; derived test pairs shareable: no; commercial use: no. TimeLapseIris agreement read; Template-Aging has its own agreement PDF of the same family (not read in full).Iris boundaries and 10 crypt/furrow points placed by two people on 30 pairs.
neurite-OASIS 2Dcoronal brain T1 slices, inter-subject414 subjects, any pair4, 24 and 35-label maps on every imageboth imagesdirectOASIS data use agreement (custom), redistribution under the same termsnon-commercialleft out
slices of the OASIS subjects already in the corpus (L2R_OASIS), labels from FreeSurfer
NHANES II cervical and lumbar spine radiographsLateral cervical (flexion) and lateral lumbar radiographs, US survey 1976-1980; between subjects, 7,377 people with both17,075 images from 9,698 participantsNone in the release; Zenodo 10223910 adds vertebra coordinates for 600 images (+ C2/S1 by two radiologists); Medical Metrics automatic corners published as statistics onlyboth imagesdirectPublic US government data (CDC FTP); Zenodo annotations CC BY 4.0openleft out
film scans whose 600 annotated images need the CDC images plus a 1 GB archive with unnamed original annotators; CSXA and BUU-LSPINE fill the same slots
Moved from NLM to CDC in 2025 (status change vs older NLM links).Use the 600 annotated images; verify a 50-image subset by a second reader
NIH ChestX-ray14 (follow-up visits)Frontal chest radiographs, NIH Clinical Center, same patient over follow-up visits112,120 images of 30,805 patients; follow-up number and patient ID per image14 NLP labels; 984 boxes on 880 single images (unverified count)one imagedirectNo restriction on use; cite paper and acknowledge NIH CC (as stated by Google Cloud mirror)opentruth to make
open longitudinal radiographs by patient and follow-up number, no account; two people place 20 landmarks on 50 pairs, CheXmask's network masks as an apart check
Open longitudinal CXR source with no account. Images are 1024x1024 8-bit PNG, no DICOM spacing. status change: Round 4 count: listed in 6 of the 38 round-4 surveys and benchmark papers (Boveiri 2020 review; Bharati 2022 review; Hammoudeh 2023 survey; Chen X 2021 review (Prog Biomed Eng); Han 2022 GAN registration review; Foundation models review 2026).Build pairs from patient ID + follow-up number; two people place 20 landmarks on 50 pairs; CheXmask covers these images for a weak mask overlap check
NLST chest radiographs (ACRIN sites)Annual screening chest radiographs (CXR arm), same participant over 3 roundsOnly films from ACRIN sites (of 18,840 ACRIN participants, both arms); LSS-site films were never digitisedScreening results; no image annotationsone imagerequestrequest to ECOG-ACRIN data sharing (CDAS does not host the radiographs); no terms publishedaskneeds a request
the data needs a request or a data use agreement
Status unclear: number of digitised films not published; request via ECOG-ACRIN. Licence checked on 2026-10-10 at https://cdas.cancer.gov/learn/nlst/images/; derived test pairs shareable: unclear; commercial use: unclear. Only ACRIN-site X-rays were digitised; LSS-site films are not available.Same as PLCO: landmarks on a 50-participant subset
Normative AOSLO montage repository (Cooper 2024)Adaptive-optics SLO confocal, split-detection and dark-field triplets montaged over 6 mm, 50 normal eyes4,039 triplets (57-111 per montage), 50 montagesMontages from automated montaging (Chen 2016 software); triplets aligned by construction; no manual registration truthone imagedirectnot visible: figshare private link does not resolve through the API or plain fetch; article CC BY 4.0, code Apache-2.0unknowntruth to make
automatic montages; simultaneous confocal and split-detection tiles are aligned by construction (known warp); the figshare link is private-share
Whether raw tiles with positions are included is unverified. Three AO channels captured simultaneously. Licence checked on 2026-10-10 at https://figshare.com/s/8263d11ac4200b9c3723; derived test pairs shareable: unclear; commercial use: unclear. Article and code licences do not cover the image data; check whether a public figshare record replaced the private link.Known warp between simultaneously captured confocal and split-detector tiles gives exact multimodal truth; montage positions as tile-overlap truth.
OAI pelvis landmarks (Zenodo 5457112)AP pelvis radiographs from OAI (Swissray scanner); between subjects524 subjectsFemoral head centre, upper acetabular rim, innermost point of subchondral sclerosis, teardrop, by a radiologist and reviewed by an orthopaedic surgeonboth imagesaccountCC BY 4.0 (coordinates); images need OAI/NDA accountopenneeds an account
the download needs an account
Only 8 landmarks per image (4 per side). OAI has repeat pelvis films for some visits (unverified).Add 15 more pelvic landmarks (ASIS, obturator foramina, pubic symphysis) on 50 images by two readers
OCTA and fundus multimodal DR dataset (Pune)colour fundus to OCTA, same eye111 pairs, 76 peoplenoneone imagedirectaccess on prior permission stating the use, with citation and DOI (Zenodo 8375220, files restricted)askneeds a request
Zenodo files restricted: access on stated purpose; no truth either
Hosting not checked. Licence checked on 2026-10-10 at https://zenodo.org/records/8375220; derived test pairs shareable: unclear; commercial use: unclear. Data in Brief 2024 (PMC10828556) points to the same Zenodo record as the 'Pune multimodal OCTA and widefield fundus' row; the article is CC BY but the data are not.vessel bifurcations found by a program and verified
OCTA-Mosaicking Dataset (A Coruna / UNED)Overlapping 6x6 mm superficial-plexus OCTA en-face images (Topcon Triton) from disc to temporal retina, same eye, for wide-field mosaicking19 sets of 3-5 images (71 images), 10 patientsNone manual; mosaic results from the authors' evolutionary method includedone imagedirectCC BY-NC 4.0non-commercialtruth to make
19 sets of overlapping OCTA images, no manual truth, non-commercial; two people place 10 capillary-junction pairs in about 50 overlaps
Real motion artefacts; small and clean. Fails expert truth.Two people place 10 capillary-junction pairs in each overlap (about 50 overlaps).
OCTA60 (CARe, OCTA to widefield fundus)12x12 mm OCTA (SVision VG200) vs widefield colour fundus (Zeiss Clarus 500), large field-of-view gap, same eye60 pairs>=10 manual keypoint correspondences per pairboth imageshiddenrelease not statedunknownleft out
its images or its truth are not released
Repo has demo only; ask Renmin Univ (Xirong Li). Used in 1 of the 2D and microscopy registration papers read (Cross-modal large-FoV fundus registration 2025).
OdontoAI Open Panoramic Radiographs (O2PR)Panoramic dental radiographs, Brazil; between subjects (replaces UFBA-UESC and DNS Panoramic sets)4,000 images, 2,000 with public labelsInstance masks for all 52 tooth types (permanent and deciduous) with numbering, human-in-the-loop verifiedboth imagesrequestsigned agreement: research only; redistribution, modification and commercial use illegal without IvisionLab's permission; no third-party accessnon-commercialneeds a request
the data needs a request or a data use agreement
Also lists TRPR (8,029 panoramics with text reports, no masks). Licence checked on 2026-10-10 at https://github.com/IvisionLab/MEDIA-datasets; derived test pairs shareable: no; commercial use: no. Request must come from an associate or full professor's institutional account with a digitally signed PDF, which a company cannot provide.
OMEGA Basel geographic atrophy FAF seriesSpectralis FAF of GA eyes, up to 4 visits 12 weeks apart37 eyes, 30 patientsGA masks by three graders (majority vote); visits pre-registered with RIFT by the authorsboth imageshiddenin-house (University Hospital Basel)askleft out
its images or its truth are not released
code public, data notvessel bifurcation landmarks per visit pair by two people
OPTIMAM Mammography Image Database (OMI-DB)UK NHS screening mammograms with prior rounds (3-yearly), several sites and vendorsunverifiedExpert lesion annotations (boxes/outlines) on cancer cases; priors labelled by round (unverified)requestper-project data sharing agreement with Cancer Research Technology (CRT owns the IP) after scientific review by an OPTIMAM sub-committeeaskneeds a request
the data needs a request or a data use agreement
Access needs project approval; no fee stated. status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (MGRegBench 2025). Licence checked on 2026-10-10 at https://medphys.royalsurrey.nhs.uk/omidb/getting-access/; derived test pairs shareable: unclear; commercial use: unclear. Commercial organisations are explicitly eligible to apply; each further project needs a new application; agreement text not public.Interval cancers: mark the lesion on prior and current by two readers
PAD-UFES-20smartphone clinical photos of skin lesions, some lesions photographed more than once2,298 images of 1,641 lesions (1,373 patients) per data descriptornone spatialone imagedirectCC BY 4.0 (Mendeley)opentruth to make
smartphone photos of skin lesions, few repeats, no truth; two people click 4-6 landmarks on the repeat pairs
Uncontrolled smartphone views; few repeats per lesiontwo people click 4-6 landmarks on repeat pairs
PadChest / PadChest-GRChest radiographs, Hospital San Juan Alicante, same patient over time; PadChest-GR adds grounded findingsPadChest: >160,000 images of about 67,000 patients; PadChest-GR: 4,555 studiesPadChest: report labels with 104 anatomical location terms (no boxes); PadChest-GR: radiologist boxes per finding (14 radiologists, up to two independent sets), single studiesone imagerequestPadChest Dataset Research Use Agreement (free for researchers, no sharing)non-commercialneeds a request
the data needs a request or a data use agreement
Gaggion landmark set and CheXmask include PadChest images.Pair repeat studies by patient ID; two people place 20 landmarks on 50 pairs; reuse Gaggion/CheXmask contours on the PadChest images that have them
Paired retina tabletop vs portable fundus (RetSyn, Hugging Face)colour fundus of the same eye from a tabletop and a smartphone-based portable camera, taken together327 patientsnone (diabetic retinopathy labels only)directMITopentruth to make
tabletop and portable photos of one eye, not fully released; bifurcations by program, a person checks
card says the full set is released after paper acceptance; 404 files present; cross-device, not longitudinalvessel bifurcations found by a program then verified by a person on 20 pairs
Panoramic Dental X-rays With Segmented Mandibles (Abdi)Panoramic radiographs, Noor Medical Imaging Center (Qom, Iran), healthy to edentulous; between subjects116 patientsMandible manually segmented by two dentistsboth imagesdirectCC BY-NC 3.0non-commercialtruth to make
mandible masks by two dentists, one structure that an affine already aligns closely; add tooth labels or landmarks
Two-reader masks allow an inter-rater floor.
PAX-Ray++ (CT-projected chest radiographs with 157 anatomy labels)Synthetic frontal and lateral chest X-rays projected from 7,377 thorax CTs; between-subject 2D registration, frontal-to-lateral via the source CT14,754 images (frontal+lateral)157 projected anatomy masks (bones, lobes, vessels) from nnU-Net pseudo-labels on CT, automaticboth imagesdirectCC BY-NC-SA 2.0 (HF API)non-commercialleft out
CT projections, not radiographs, labelled by a network; Aariz and CSXA cover radiographs with expert truth
Labels are automatic; images are DRR-like, not real X-rays. The smaller PAX-Ray (852 images) has the same design.Warp the source CT with a known smooth field and re-project: exact 2D truth for intra-subject pairs.
PKU cephalogram datasetLateral cephalograms; inter-subject102 images19 landmarks by cliniciansboth imagesdirectCC BY-SA 4.0 (per CL-Detection paper)openneeds an account
102 cephalograms distributed with CL-Detection 2023 on grand-challenge (account and agreement); Aariz covers the problem openly
Host not checked.
Plantar Thermogram Database (Hernandez-Contreras 2019)plantar thermal images, left vs mirrored right foot of the same subject; between subjects167 subjects (122 diabetic, 45 control), left and right foot eachangiosome region crops (4 per foot) onlyboth imagesdirectCreative Commons Attributionopenneeds an account
IEEE DataPort: 'login to access dataset files' even for open access; no landmarks either
Temperature CSVs per pixel; foot outline is the main structure Licence checked on 2026-10-10 at https://api.datacite.org/dois/10.21227/tm4t-9n15; derived test pairs shareable: yes; commercial use: yes. IEEE DataPort open access; version of CC BY not given in DataCite.two people click toe tips, heel and metatarsal heads on 30 left/right pairs
PLCO chest radiographsAnnual screening PA chest radiographs (up to 4 rounds), same participant over yearsAbout 198,000 digitised images from 237,000 radiographs; about 58,000 participants; requests capped at a 25,000-participant subset (~89,000 images, 850 GB)Screening results and cancer outcomes; no image annotationsone imagerequestapproved CDAS project with a fully executed Data Transfer and Use Agreement (DTUA)askneeds a request
the data needs a request or a data use agreement
Digitised film (TIF, low-contrast compression). Regular yearly intervals make clean temporal pairs. Licence checked on 2026-10-10 at https://cdas.cancer.gov/learn/plco/images/; derived test pairs shareable: unclear; commercial use: unclear. Standard 25k selection (~89,000 images, 850 GB); DTUA text not read.Pick 50 participants with 3-4 rounds; two people place 20 landmarks per image; nodules confirmed later give lesion points on some pairs
PRIME-FP20Ultra-widefield fundus photo vs FA, same visit15 pairsManual vessel maps on the fundus photosone imageaccountCreative Commons Attributionopenneeds an account
the download needs an account
Small; truth on one side. Used in 1 of the 2D and microscopy registration papers read (Cross-modal large-FoV 2025 (training)). Licence checked on 2026-10-10 at https://api.datacite.org/dois/10.21227/ctgj-1367; derived test pairs shareable: yes; commercial use: yes. IEEE DataPort open access; free account to download.Draw vessel maps on the FA too, or two people click 15 bifurcations on both images, for all 15 pairs.
Proxima A/B and lampalizumab trials FAF and IR (Roche/Genentech)FAF and infrared images of geographic atrophy every 6 months over 2 years, same eye; FAF-IR same session597+ patients, 4 visits eachGA lesion annotations by graders G1/G2 at each visitboth imagesrequestclinical trial data sharing request (route unverified, likely Vivli)askneeds a request
the data needs a request or a data use agreement
FAF and IR from one session are aligned by the devicevessel bifurcations between visits by program plus person; FAF-IR same-session pairs give device alignment
Pune multimodal OCTA and widefield fundus for DRWidefield confocal SLO colour fundus (Eidon) vs macular OCTA (Optovue), same eye111 eye pairs, 76 subjectsDR class by two eye specialists onlyone imagerequestaccess on prior permission stating the use, with citation and DOIaskneeds a request
the data needs a request or a data use agreement
Large scale gap (110 deg fundus vs 8 mm OCTA). Fails truth. Licence checked on 2026-10-10 at https://zenodo.org/records/8375220; derived test pairs shareable: unclear; commercial use: unclear. Zenodo files restricted; no licence named on the record.Two people place 6 bifurcation pairs per pair on 30 pairs; vessel masks as weak check.
RA2-DREAM Challenge radiographsHand/wrist and foot radiographs of rheumatoid arthritis patients (CLEAR and TETRAD studies); some patients with more than one set674 radiographic sets from 562 patients (367 train with scores)Expert Sharp/van der Heijde scores per joint; no landmarksone imagecontrolledSynapse conditions of use (challenge closed 2020)askneeds a request
the data needs a request or a data use agreement
674 sets for 562 patients, so about 100 or more patients have a second set (possible longitudinal hand pairs); check dates.Two people place 37 joint landmarks per hand on the repeat sets; joint-level score changes flag erosions
Rad lower limb X-ray dataset (Tehran)Bilateral weight-bearing full-leg radiographs (DICOM, about 0.1 mm), with osteotomy, TKA, THA and untreated cases; between subjects426 subjects (419 annotation files)26 landmarks (13 per leg)both imagesdirectnone stated; page gives no licence or access routeunknownneeds a request
26 landmarks on full-leg radiographs would pass, but the page gives no download route: ask
Access mode not stated on page. Licence checked on 2026-10-10 at https://aailab.ir/datasets/xray_rad_lower_limb.html; derived test pairs shareable: unclear; commercial use: unclear. Dataset description only; no download link or terms on the page.
RAM-H1200Full-hand radiographs of rheumatoid arthritis patients from 6 centres; between subjects1,200 hand radiographsWhole-hand bone instance segmentation, pixel-level erosion masks, SvdH joint regions and joint scoresboth imagesdirectCC BY-NC-SA 4.0non-commercialneeds an account
per-bone instance masks of 1,200 hand radiographs would give dense correspondences, but the Hugging Face set is gated (account and terms), non-commercial
Per-bone instance labels give dense inter-subject correspondences (bone k to bone k); joint boxes as points.
RECOVERY-FA19 (UWF FA vessel detection)Ultra-widefield fluorescein angiography (Optos California P200DTx), single images8 images, no pairsManual binary vessel mapsone imageaccountCreative Commons Attributionopenneeds an account
the download needs an account
Same group as FLoRI21 and PRIME-FP20. Fails pair criterion. Licence checked on 2026-10-10 at https://api.datacite.org/dois/10.21227/m9yw-xs04; derived test pairs shareable: yes; commercial use: yes. IEEE DataPort open access; free account to download.Known smooth warp plus FA phase intensity change applied to each image; warped vessel maps as exact truth.
RegGR1945 (historical aerial 1945 to satellite 2015/2025, Greece)Historical aerial photo (1945) vs satellite image (2015/2025), Greece50 pairs (2015) + 30 pairs (2025, satellite images not included)Ground-truth registration coordinates (translation only)both imagesdirectother-openopentruth to make
translation-only truth, which an alignment removes; add 10 manual point pairs a pair (non-medical, belongs with the remote-sensing group)
Non-medical; translation-onlyAdd 10 manual point pairs per image to test non-rigid residuals
Repeated Dermoscopic Images of Melanocytic Lesions (ISIC)dermoscopy of the same torso nevus taken five times in sequence (same visit), some 1-year follow-up, 66 patients117 series of 5 images (585 images) of 116 lesionsnone spatial (lesion IDs only)one imagedirectCC BY-NC (per image)non-commercialtruth to make
5 repeats a nevus in one visit and some 1-year follow-ups, non-commercial; two people click dots and globules on 30 pairs, follow-ups first
Small rotations/shifts and lighting changes between repeats: a clean same-lesion rigid-to-mild deformation settwo people click 6-10 dots/globules on 30 pairs; or lesion masks as weak check
Retinal DR longitudinal fundus pairs (Tianjin, Hugging Face)colour fundus photos at baseline and 2-year follow-up, diabetic retinopathy, same eye1,148 usable pairs from 574 patientsnone (DR progression labels only)one imagedirectCC BY-NC 4.0non-commercialtruth to make
1,148 baseline and 2-year pairs with labels only, uploaded by an individual (source unchecked), non-commercial; 10 bifurcation pairs on 30 pairs
Uploaded by an individual; original source paper not checked. Different cameras/resolutions at the two visits.two people place 10 vessel bifurcation pairs on 30 pairs
Retinal IPA in-house OCT-SLO pairs (Vanderbilt)OCT en-face vs SLO, human, same eye18 pairs (+228 mouse unlabelled)8-12 keypoints per image by two annotatorsboth imageshiddenin-houseunknownleft out
its images or its truth are not released
Not released.
ReXGradient-160KChest radiographs with reports, 79 sites in 3 US health systems, some patients with several studies160,000 studies, 273,004 images, 109,487 patientsReports onlyone imageaccountReXGradient-160K Non-Commercial Data Access and Use Agreement v1 (2025)non-commercialneeds an account
the download needs an account
Gated HF repo: accept agreement. Multi-site and many devices, good for domain shift.Pair repeat studies; two people place 20 landmarks on 50 pairs
RHPE (Radiological Hand Pose Estimation)Hand radiographs of children and adolescents (Colombia, Fundacion Santa Fe de Bogota), left and right hands; between subjects6,288 images (5,492 train, 716 val, 80 test)17 keypoints and hand box per image, at least 4 trained annotators each with outlier rejection; bone age by 2 radiologistsboth imagesdirectUniversidad de los Andes / Fundacion Santa Fe de Bogota agreement: non-commercial research and educational use onlyaskin the corpuspriority 3
In the corpus: the provider's 79 test images paired between children on the 17 keypoints (39 test pairs), 30 development pairs from the validation split; research-only terms that limit sharing (ask).
Test annotations hidden behind eval.ai server. Licence checked on 2026-10-10 at https://drive.google.com/file/d/1GcgGn5nbPcyOmgXp_2CWzjLhxCT9fAQf/view; derived test pairs shareable: yes; commercial use: no. Terms linked from https://bcv-uniandes.github.io/baar-wp/; colleagues may get access if they agree to the same terms; a for-profit employer is bound too.Expert review of keypoints on 100 images; add 20 more bone-centre points per hand for denser truth
RIDIRP (retinal images of infants and ROP)Infant fundus (RetCam 3, RetCam Envision, Phoenix ICON), repeated screening sessions, same eye6,004 images, 188 infants; many eyes with 2-9 sessionsClinical/ROP labels, no correspondences (COph100 is the annotated subset)one imagedirectCC BY 4.0 on the figshare collection (the paper's code availability section says CC0)opentruth to make
the parent of COph100: its other eyes need the same 10-point protocol, and COph100's images stay with COph100
Parent of COph100; extra eyes beyond the 100 annotated ones. Licence checked on 2026-10-10 at https://doi.org/10.6084/m9.figshare.c.6626162.v1; derived test pairs shareable: yes; commercial use: yes. Both statements are open licences; DataCite rights for the figshare collection read CC BY 4.0, so credit the authors.Extend COph100 protocol: 10 manual bifurcation pairs on more eyes, or vessel masks from a public segmenter as weak check.
RIM-ONE r3 (stereo fundus)Stereo optic-disc fundus image pairs (side by side), glaucoma screening, same eye same instant159 stereo pairsDisc and cup masks by two experts, on the left image onlyone imagedirectnone stated on the original MIAG-ULL downloads pageunknowntruth to make
stereo disc pairs with masks on the left image only; 10 point pairs around the disc on 30 pairs; no licence stated
Stereo disparity, not temporal; truth on one side. Fails truth. Licence checked on 2026-10-10 at http://web.archive.org/web/20230602153401/http://medimrg.webs.ull.es/research/downloads/; derived test pairs shareable: unclear; commercial use: unclear. Idiap bob page unreachable; archived ULL page links r1-r3 with no terms; the newer RIM-ONE DL repo has no recognised licence.Two people place 10 point pairs around the disc on 30 stereo pairs; or OCT-free disparity check by vessel crossings.
RISeR-real (retinal image sequence registration, real)Longitudinal colour fundus sequences, myopia onset/progression, same eye10 sequences of 5-6 images (~50 frame pairs)12 manually annotated control-point sets per sequence (annotator not stated)both imagesaccountCC BY 4.0openneeds an account
the download needs an account
IEEE DataPort subscription required; groupwise/sequence registration. Licence checked on 2026-10-10 at https://api.datacite.org/dois/10.21227/xpe1-e969; derived test pairs shareable: yes; commercial use: yes. Licence is open but the download is an IEEE DataPort standard dataset (subscription).
RISeR-syn (retinal image sequence registration, synthetic)Synthetic fundus sequences generated from 10 real images of different eyes50 sequences of 6 framesDense control points known by construction, incl. inter-frame overlap pointsboth imagesaccountCC BY 4.0openneeds an account
the download needs an account
Synthetic warps of real fundus; IEEE DataPort subscription. Licence checked on 2026-10-10 at https://api.datacite.org/dois/10.21227/qcs2-aw02; derived test pairs shareable: yes; commercial use: yes. Licence is open but the download needs an IEEE DataPort subscription.
ROCO and MedICaT (figure-caption sets)figures from open-access papers with captionsnonecaptionsone imagedirectMedICaT: non-commercial use only, research only, plus each source article's licence; ROCO: no licence, figures keep their PubMed Central open-access article licencesnon-commercialleft out
figures from papers, no pairs
Fails pairs: one image per subject, no repeat or second modality. Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Foundation models review 2026). Licence checked on 2026-10-10 at https://github.com/allenai/medicat; derived test pairs shareable: unclear; commercial use: no. MedICaT code is Apache-2.0 but the dataset is non-commercial; ROCO images could be filtered to CC BY/CC0 articles per item.Figures without pairs: no usable truth.
RODREP longitudinal fundusColour fundus, 70 diabetic patients x 2 exams; same eye~140 eye pairsNonedirectRotterdam Ophthalmic Institute licence agreement: free for scientific research, no commercial activities, cite; alternative licence on requestnon-commercialneeds a request
rodrep.com answers 404 since 2025; ask the Rotterdam Ophthalmic Institute; no correspondences either
Fails truth (COph100 authors excluded it for that reason). Licence checked on 2026-10-10 at http://web.archive.org/web/20250519075301/http://www.rodrep.com/longitudinal-diabetic-retinopathy-screening---description.html; derived test pairs shareable: unclear; commercial use: no. Live rodrep.com page now 404; archived 2025 page read. Full agreement text ships with the download. Institute invites contact for other uses.Two people place 10 vessel bifurcations on both exams of 20 eye pairs; AutoMorph vessel junctions verified by a person speed it up.
RSNA Pediatric Bone Age + BAAR keypointsPediatric left-hand radiographs (Stanford, Colorado); between subjects of different ages14,236 images (12,611 train, 1,425 val, 200 test)Bone age by experts; BAAR adds hand boxes and 17 anatomical keypoints (phalanges, carpals, radius, ulna) from multiple trained annotatorsboth imagesdirectRSNA images: academic and non-commercial use with attribution; BAAR annotations: terms on site (not named)non-commercialleft out
the hand slot is the Digital Hand Atlas (fallback RHPE); BAAR points by trained annotators, RSNA images under their own terms
Images from RSNA challenge page; keypoints from BAAR. Growth means shapes differ strongly across ages, a hard inter-subject case.Keypoints are crowd-trained, not expert: have one expert check 100 images; for temporal pairs, pair images of similar bone age or use the Digital Hand Atlas
Schroth exercise RCT radiographs for AIS (Alberta, Borealis)scoliosis radiographs at baseline, 6 and 12 months, plus surface topography, same adolescentstrial cohort (size unverified), 3 radiographs eachCobb angle and apical rotation measurementsboth imagesrequestunverified (Borealis study record)askneeds a request
the data needs a request or a data use agreement
record lists radiographic data; image release unverifiedvertebral corner landmarks by two people on 10 patients
SCR (Segmentation in Chest Radiographs)JSRT chest radiographs; inter-subject247 imagesManual lung, heart and clavicle masks (2 observers)both imagesdirectfree for research and educational purposes; redistribution and commercial use prohibitednon-commercialneeds a request
the SCR host is gone and the JSRT images are given on application; non-commercial, no redistribution
URL returned 404 on 2026-10-10; find the new host. Licence checked on 2026-10-10 at http://web.archive.org/web/20110301093521/http://www.isi.uu.nl:80/Research/Databases/SCR/; derived test pairs shareable: no; commercial use: no. Live page gone; archived ISI page read. Images are JSRT radiographs, also non-commercial and non-distributable.
Shenzhen CXRPA chest radiographs, TB screening; inter-subject662 images (566 with masks)Lung masks added later by a third party (Stirenko et al.)both imagesdirectnone stated; NLM says it made the set publicly available for researchunknownleft out
the same inter-patient chest problem as Montgomery, masks by a third party; one chest set is enough, and lungs alone leave no room after an affine
status change: Round 4 count: listed in 3 of the 38 round-4 surveys and benchmark papers (Hammoudeh 2023 survey; Chen X 2021 review (Prog Biomed Eng); Han 2022 GAN registration review). Licence checked on 2026-10-10 at https://pmc.ncbi.nlm.nih.gov/articles/PMC4256233/; derived test pairs shareable: unclear; commercial use: unclear. Jaeger et al. QIMS 2014 (the NLM release paper) has no licence or terms; Open-i FAQ page now only shows a funding-lapse notice.
SIGF (Sequential fundus Images for Glaucoma Forecast)Longitudinal optic-nerve-head colour fundus, same eye, 1986-2018405 sequences, 3,671 images (avg 9 per eye); 37 progressing sequencesGlaucoma labels only; EyeLiner authors annotated 10 clinician keypoints on 37 first/last pairs (not released, unverified)one imagerequestno licence; available again on request (xw405@cam.ac.uk) after an ethics reviewaskneeds a request
the data needs a request or a data use agreement
Access changed: open, then withdrawn, now on request (README 2025). Licence checked on 2026-10-10 at https://github.com/XiaofeiWang2018/DeepGF; derived test pairs shareable: unclear; commercial use: unclear. README: open access was withdrawn, then reopened on request; no terms published.Ask QTIM-Lab for EyeLiner's 37-pair keypoints; else two people place 10 bifurcation points on 40 pairs.
SLICE-3D (ISIC 2024 challenge)15x15 mm lesion crops from 3D total-body photography (Vectra WB360), patient ID and 3D lesion coordinates401,059 crops; one crop per lesion, single time pointnone spatial for pairs; 3D coordinates of each lesion on the bodyone imagedirectCC BY-NC 4.0non-commercialleft out
one crop a lesion, no repeat capture
Fails pairs: no repeat captures; 3D coordinates could support constellation matching onlynone practical without repeat TBP captures
SLIMBRAIN (intraoperative hyperspectral, RGB and depth of brain surface)hyperspectral, RGB, depth and multi-view images of the exposed brain during surgery, same patient193 patientsneurosurgeon labels on hyperspectral pixels; camera alignment not givenone imagedirectunverified (e-cienciaDatos)unknownneeds a request
access after a signed data use agreement e-mailed to the providers; no camera-to-camera truth either
Sci Data 2025.two people place 10 vessel landmarks on HS and RGB for 20 patients
spatialLIBD DLPFC Visium (Maynard 2021)Human dorsolateral prefrontal cortex Visium: 12 sections from 3 donors, each donor two pairs of adjacent 10 µm sections 300 µm apart; same subject, serial sectionsPer donor: 2 adjacent pairs and 300 µm pairs (used by PASTE, PASTE2, STalign, SLAT)Manual cortical layer (L1-L6, WM) labels on all spots of all 12 sectionsboth imagesdirect: spatial-dlpfc.s3.us-east-2.amazonaws.com (no account), Globus as an alternativeArtistic-2.0 (spatialLIBD Bioconductor package; LieberInstitute/HumanPilot and spatialLIBD repositories); the AWS images carry no separate noticeaskin the corpuspriority 3
In the corpus: the six adjacent pairs of the three donors, four test pairs of two donors and two development pairs of the third, the spots' layers as label maps; ranked apart (two test donors, 100 µm grain).
Small, but the standard spatial-omics slice-alignment benchmark. Used in 5 of the 2D and microscopy registration papers read (SLAT 2023; SPACEL 2023; Hu 2024 benchmark; Spotscape 2025; ...). Licence checked on 2026-10-10 at https://bioconductor.org/packages/release/data/experiment/html/spatialLIBD.html; derived test pairs shareable: yes; commercial use: yes. Artistic-2.0 is declared for the experiment-data package that ships the Visium objects; the full-resolution TIFFs on AWS carry no separate notice.
Spondylolisthesis Vertebral Landmark (Mendeley)Lateral lumbar radiographs (208 Honduran + 508 from BUU-LSPINE); between subjects716 images, 698 annotated4 corner keypoints and a box per vertebra L3-S1both imagesdirectCC BY 4.0openleft out
508 of its 716 images are BUU-LSPINE's; 16 corner points, annotator unnamed
Overlaps BUU-LSPINE; Honduran part is new.
SuperRetina Lab keypoint setColour fundus with manual keypoint labels (single images), plus unlabelled Auxiliary setsingle images (counts unverified; Lab-Aux 919 images per GeoFormer)Manual keypointsone imagehiddennot released (sample file only)unknownleft out
its images or its truth are not released
Repo also notes FIRE pair P37 ground truth is wrong (exclude it).Not released: no truth possible.
Syn3DWoundrendered multi-view photos of synthetic wounds on 3D human avatars (Blender Cycles), camera trajectories around each woundmany views per wound (count unverified); CSIRO data portal collection csiro:73258camera intrinsics/extrinsics, meshes and masks: exact multi-view correspondenceboth imagesdirectnone stated (project page and GitHub repo have no licence)unknownleft out
renders of synthetic wounds on licensed avatars (RenderPeople, 3DBodyTex), no licence
Synthetic, exact truth; good for testing photo registration of a curved skin surface Licence checked on 2026-10-10 at https://github.com/lebrat/Syn3DWound; derived test pairs shareable: unclear; commercial use: unclear. Avatars come from RenderPeople and 3DBodyTex, both under restrictive licences, which may bind the renders.
syntheticCAMUSsimulated 2D echo sequences100known dense motion field (synthetic)both imagesdirectnone stated for syntheticCAMUS; built from CAMUS, which is CC BY-NC-SA 4.0 non-commercial research onlyaskleft out
its motion fields are public training data, and CAMUS gives expert truth for the same problem
Used in 2 of the 2D and microscopy registration papers read (EchoDiST 2026; EchoTracker2 2026). Licence checked on 2026-10-10 at https://humanheart-project.creatis.insa-lyon.fr/database/api/v1/collection?limit=50; derived test pairs shareable: unclear; commercial use: unclear. No LICENSE file in the syntheticCAMUS collection, unlike CAMUS and TED; likely inherits CAMUS NC-SA, confirm with CREATIS.
SZCH-X-Rays (dual-energy soft-tissue pairs)PA chest radiograph and dual-energy subtraction soft-tissue image of the same exposure (GE Discovery XR656), Shenzhen741 pairs (592/74/75); only the 75 test pairs released so farSame acquisition, near-aligned by construction (two-shot DES, small cardiac motion possible)both imagesdirectCC BY-NC 4.0non-commercialtruth to make
radiograph and soft-tissue image of one exposure, near-aligned; a known warp gives the truth; only 75 test pairs out, non-commercial
Training/validation sets promised after paper acceptance. 2021x2021 px, 0.194 mm.Known smooth warp of one channel gives exact cross-appearance truth; check residual DES motion first
TED2D echo over the full cycle, same patient98manual LV and myocardium on every frameboth imagesdirectCC BY-NC-SA 4.0, strictly non-commercial scientific research, mandatory citationnon-commercialleft out
its 98 patients are CAMUS patients (no image twice); CAMUS's ED and ES frames carry the same kind of expert masks
Licence checked on 2026-10-10 at https://humanheart-project.creatis.insa-lyon.fr/database/api/v1/item/66e2bf49961576b1bad5094e/download; derived test pairs shareable: yes; commercial use: no. LICENSE_TERMS.md in the TED Girder collection; ShareAlike on derived pairs.
TrackRAD20252D sagittal cine MR on MR-linac, frame to frame50 labelled casesmanual target on every frameboth imagesdirectCC BY-NCnon-commercialin the corpuspriority 2
Built: TrackRAD2025 (and TrackRAD2025_Seg).
Tufts Dental DatabasePanoramic dental radiographs, Tufts School of Dental Medicine 2014-2016; between subjects1,000 panoramic imagesTeeth masks with numbering and abnormality masks by an expert and a student; eye-tracking gaze mapsboth imagesrequestTufts Dental Database release agreement: non-commercial research and educational use only; no access for colleagues; no reproduction or posting to any storagenon-commercialneeds a request
the data needs a request or a data use agreement
Per-tooth labels give inter-subject correspondences (tooth 36 to tooth 36). Licence checked on 2026-10-10 at https://docs.google.com/forms/d/e/1FAIpQLSdZOTw04C27AueKiDu224WwXNd7OSZSlZg2Sj44vkQvD8OAtQ/viewform; derived test pairs shareable: no; commercial use: no. Each user must sign the Google Form agreement; a private copy shared with the team is explicitly ruled out.
UCSD Jacobs Retina Center multimodal pairs (CF-IR, CF-FA, FAF-MC)Colour fundus vs infrared reflectance (Topcon/Heidelberg), Canon CF vs FA, Optos FAF vs multicolor, same eye530+253 CF-IR, 29 CF-FA, 319 FAF-MCPoint correspondences on CF-IR training pairs; evaluation by vessel Diceboth imageshiddenprivateunknownleft out
its images or its truth are not released
Not released; ask Nguyen/Bartsch group. Used in 1 of the 2D and microscopy registration papers read (Modality-invariant coarse-to-fine 2026).
UK Biobank retinal imaging (baseline and repeat visit)Colour fundus and macular OCT (Topcon 3D OCT-1000) at baseline and repeat assessment, same eyerepeat-visit subset (size unverified, order of 10^4 eyes)None for registrationone imagecontrolledUK Biobank access agreement (fee)askneeds a request
the data needs a request or a data use agreement
Covered generically by the existing 'UK Biobank, NAKO, raw OAI' entry; listed here for the retina track. Counts from memory, unverified.Fundus-to-OCT projection aligned by device; bifurcation points by a program verified by a person on 50 repeat pairs.
Ultrasound elastography phantom and in vivo RF data (Concordia/Rivaz)real ultrasound RF frame pairs of a CIRS breast phantom under compression; sample in vivo liver ablation RF data with the GLUE/AM2D code2,200 phantom RF pairs; a few in vivo liver sequencesnone (inclusion geometry of the phantom known)one imagedirectnone; page asks to cite the papers (site footer: Copyright 2023 IMPACT Laboratory, all rights reserved)unknowntruth to make
real RF frames without displacement truth; a known warp drawn by the corpus, unpublished, is the fair way to use the lab's data; all rights reserved, ask
No displacement truth; in vivo RF at users.encs.concordia.ca/~hrivaz/Ultrasound_Elastography/ Licence checked on 2026-10-10 at https://users.encs.concordia.ca/~impact/ultrasound-elastography-dataset-for-unsupervised-training/; derived test pairs shareable: unclear; commercial use: unclear. Direct download, no licence or terms of use for the data.known smooth warp of real RF frames, or strain-image contrast of the known inclusions as an indirect check
Ultrasound elastography simulation database (Concordia IMPACT)simulated ultrasound RF/B-mode of compressed phantoms with hard inclusions, pre vs post compression2,400 images (24 phantoms x 10 strains x 10 scatterer draws), any pair within a phantomFEM axial and lateral displacement relative to the pre-compression imageboth imagesdirectnone; page asks to cite the papers (site footer: Copyright 2023 IMPACT Laboratory, all rights reserved)unknownleft out
its FEM fields are public training data for displacement networks (MPWC-Net and others): a test on them favours those methods; all rights reserved
Exact dense displacement; small strains (0-4.5%); RF-domain problem Licence checked on 2026-10-10 at https://users.encs.concordia.ca/~impact/ultrasound-elastography-simulation-database/; derived test pairs shareable: unclear; commercial use: unclear. Direct download, no licence or terms of use for the data.
Unity Imaging echocardiography datasets2D echo frames (PLAX, apical 4 and 2 chamber), UK multi-hospital, frames from the same loop>7,500 labelled images; PLAX validation set of 100 loops with ED and ES framesexpert keypoints (PLAX LV measurements; apical LV/LA/myocardium contours as point lists), up to 13 experts twice each on validationboth imagesdirectCC BY-NC-ND 4.0 (model development dataset)non-commercialleft out
ND licence, and its paired ED/ES set is PLAX with a handful of measurement points; CAMUS covers 2D echo with dense masks
ND licence forbids sharing modified copies; multi-expert keypoints give inter-observer error on ED/ES pairs
UQ longitudinal total-body photography and dermoscopy (Ghahari 2025)dermoscopic images and lesion tiles from 3D total-body photography, same lesion at 2-7 visits480 participants (340 longitudinal); 7,038 lesions with longitudinal dermoscopy (35,909 images); 250,162 lesion tileslesion identity across visits (no point truth)direct'Other': UQ General Usage Terms and Conditions for data publishing in eSpace (data reuse agreement); access listed as openasktruth to make
the same lesion at 2-7 visits (35,909 dermoscopic images), lesion identity only; two people mark 5-10 skin features on 50 pairs; UQ terms to read
Tiles are crops from 3D avatars; whole avatars not included (unverified). Licence checked on 2026-10-10 at https://researchdata.edu.au/a-longitudinal-dataset-skin-cancers/3750284; derived test pairs shareable: unclear; commercial use: unclear. Rights from DataCite 10.48610/a13deaf; both UQ terms pages linked from the record return 404, so the clauses were not read.two people mark 5-10 skin features per dermoscopic pair on 50 pairs; lesion masks on both (weak)
UT-FSOCTA (UTHealth fundus and en-face OCTA)45 deg colour fundus (OptoVue iVue) vs superficial-plexus en-face OCTA (AngioVue) of disc and macula, same eye112 subjects, disc and macula OCTA per subjectFundus and OCTA manually aligned by the authors (cropped fundus FOVs matching each OCTA); ROSE vessel mapsboth imagesdirectnone stated (Zenodo record open, no licence field)unknowntruth to make
fundus and OCTA of 112 subjects manually aligned by the authors (crop position = a provider transform); check it with 6 bifurcation pairs on 20 subjects to reach a main board; no licence
Crop position in the full fundus gives a reference transform (manual quality). Licence checked on 2026-10-10 at https://zenodo.org/records/6476639; derived test pairs shareable: unclear; commercial use: unclear. Files public on Zenodo, but no licence or terms in the record.Check the manual alignment with 6 bifurcation pairs on 20 subjects.
Vanderbilt SECTR 9-field OCT en-face mosaicking setEn-face projections of OCT sub-field volumes (SECTR system) for mosaicking; same eye, overlapping fields2 subjects x 9 sub-fields, 144 registration pairsNoneone imagehiddenNot statedunknownleft out
its images or its truth are not released
Private (IRB 202230); fails the truth criterionVessel bifurcations found by a program then verified by a person in overlap zones; or tile OCTA-500 en-face images with known warps
VARIAFundus images, same person at different times (biometrics)unverifiedIdentity labels onlyrequestpassword by e-mail (mortega@udc.es), stated as for statistical purposes only, no fee; citeaskneeds a request
the data needs a request or a data use agreement
Fails truth. Licence checked on 2026-10-10 at http://web.archive.org/web/20260518065800/http://www.varpa.es/research/biometrics.html; derived test pairs shareable: unclear; commercial use: unclear. Live VARPA page now 404; archived copy from 2026-05 read. No licence named.Check how many eyes have repeats; two people place 10 vessel bifurcations on 20 same-eye pairs.
VinDr-CXRPA chest radiographs, two Vietnamese hospitals; between subjects only18,000 images (15,000 train, 3,000 test)Finding boxes by 3 radiologists (train) or 5-radiologist consensus (test); no anatomy landmarksone imagecontrolledPhysioNet Credentialed Health Data License 1.5.0askneeds a request
the data needs a request or a data use agreement
Fails: no patient IDs, no repeat studies; pathology boxes are not correspondences.Inter-subject only: run a contour landmark model (HybridGNet) and verify on 50 images
VinDr-MammoFour-view screening mammograms, two Vietnamese hospitals; between views and between subjects5,000 exams, 20,000 imagesBI-RADS 3-5 finding boxes, double reading with arbitration by a third radiologistone imagecontrolledPhysioNet Restricted Health Data License 1.5.0askneeds a request
the data needs a request or a data use agreement
No priors. Left/right symmetry pairs (contralateral registration) possible. status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (MGRegBench 2025).Contralateral pairs: two readers mark nipple and pectoral line points; boxes on CC/MLO for lesion matching
VinDr-RibCXRFrontal chest radiographs; between subjects245 images (196 train, 49 validation)20 individual ribs outlined and labelled by human expertsboth imagesrequestVinDr-RibCXR Data Use Agreement: non-commercial research only; no distribution or reproduction to third parties; no modification; each user registers individuallynon-commercialneeds a request
the data needs a request or a data use agreement
Labelled ribs give dense inter-subject correspondences (rib k to rib k), ideal to measure rib misalignment. Licence checked on 2026-10-10 at https://drive.google.com/file/d/1Wr3iI7-OwZHD4eWtCALpRZKvhJuemKAs/view; derived test pairs shareable: no; commercial use: no. DUA linked from https://github.com/vinbigdata-medical/MIDL2021-VinDr-RibCXR (code MIT).
VinDr-SpineXRSpine radiographs, Vietnamese hospitals; between subjects10,466 images, 5,000 studiesLesion boxes (13 types) by one of three radiologists; no landmarksone imagecontrolledPhysioNet Restricted Health Data License 1.5.0askneeds a request
the data needs a request or a data use agreement
Fails: no anatomical correspondences; several views per study.Vertebral corner landmarks by a public model, verified on 50 images
WoundsDB (chronic wounds multimodal image database)chronic wound photo vs thermal image vs depth/stereo (3D mesh), same wound, parallel camera axes; repeated visits188 image sets from 79 patient visits (188 RGB, 188 thermal, 184 stereo, 177 depth)authors' co-registration parameters and transformed images; expert wound outlines by a surgeonboth imagesdirectnone: the site's Terms of use page says 'under construction'; download needs an accountunknownneeds an account
download needs an account and the site's certificate is broken; its RGB-thermal registration is the authors' rig calibration
Rare RGB-thermal-depth multimodal truth; registration parameters are the authors' method, not independent; site TLS broken on 2026-10-10 Licence checked on 2026-10-10 at https://chronicwounddatabase.eu/Terms; derived test pairs shareable: unclear; commercial use: unclear. Read with TLS verification off (certificate still broken); Silesian University of Technology, cite Access 2020 paper.two people click 6-10 wound-border and skin features on RGB and thermal for 20 sets; calibration target images if available
X-Ray Angiography Images and SYNTAX-Score DatasetCoronary angiogram frames with SYNTAX scores; several views per patient2,943 angiograms from 231 patients (XA-170K description)SYNTAX score per patient; no spatial marks (unverified)one imagedirectCC BY 4.0openleft out
single frames of several views with SYNTAX scores: no pair, no spatial truth
Fails: single frames, no correspondences.Segment labels on two views of the same artery for multi-view (not registration) checks
XCAD (X-ray coronary angiography, SSVS)Coronary angiograms with matching pre-contrast background frames (used for self-supervised vessel extraction)1,621 training angiograms plus background frames; 126 test images with masksVessel masks on 126 test images by expertsone imagedirectnone (no licence file; Dropbox/Baidu links in the README, cite the paper)unknowntruth to make
background and contrast frames of the same runs, masks on test images only; landmarks on background frames by two people; no licence
Background frames are from the same runs before contrast (DSA mask/live style); pairing per run unverified. Licence checked on 2026-10-10 at https://github.com/AISIGSJTU/SSVS; derived test pairs shareable: unclear; commercial use: unclear. README only asks for a citation.Pair background and contrast frames of the same run; background landmarks by two people on 30 pairs

2D/3D · 112

DatasetProblemPairsChecked againstAccessLicenceStatusNotesTo make the truth
10-case CT and fluoroscopy 2D-3D set (Science Data Bank)CT volume and several fluoroscopy images per case10 casesnot stateddirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Chinese Science Data Bank; DICOM only.two people mark fiducials or bony points in CT and each X-ray; solve pose
1000 Lungs (SynAirG synthetic bronchoscopy)synthetic chest CTs (MAISI) with airway meshes and rendered bronchoscope episodes (depth, normals), video-to-CT1,000 episodesexact camera trajectory (TUM poses) and mesh-to-CT registration per caseboth imagesdirectOpenMDW-1.1openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Fully synthetic anatomy; RGB not yet rendered in geometry-first batches Licence checked on 2026-10-10 at https://huggingface.co/datasets/chrisvoncsefalvay/1000lungs; derived test pairs shareable: yes; commercial use: yes. HF card licence field; OpenMDW is a permissive model/data licence (attribution, notice)
2DeteCT (2D experimental fan-beam CT)Experimental 2D fan-beam sinograms and reconstructions of 5,000 slices of mixed objects (nuts, fruit, rocks)5,000+ slicesExact fan-beam geometry; reference reconstructionsboth imagesdirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Fails the 3D criterion: 2D slices only (1D projections to 2D image); non-medicalOnly 1D-to-2D toy registration by construction
ABBA demo: serial coronal sections of an adult mouse brainSlide-scanner fluorescence (DAPI, autofluorescence, mCherry) of 20 µm sections every 80 µm; section-to-atlas (Allen CCF)One series of sections to the atlasNoneone imagedirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Test data for the ABBA Fiji/QuickNII-style workflow; newer version exists. Licence checked on 2026-10-10 at https://zenodo.org/records/6553641; derived test pairs shareable: yes; commercial use: yes. Zenodo record licence fieldTwo people align with QuickNII/VisuAlign (as DeepSlice did) and add 10 landmarks per section; inter-rater spread sets the bar
ABC Atlas MERFISH whole mouse brain sections mapped to CCFv3 (Allen, Yao 2023)59 serial coronal MERFISH sections (200 µm apart, ~4.3 M cells, 500 genes) of one mouse vs Allen CCFv3; section-to-volume59 sections to one atlasAllen global affine plus per-section deformable mapping (point-to-point coordinates); midlines set manuallyone imagedirectCC BY-NC 4.0non-commercialwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Cell-type labels per cell give a label check against CCF regions; images released only downsampled.Cell-type-defined regions (e.g. layer-specific types) as labels on both section and atlas; two people place 10 landmarks per section on the DAPI-like image
Allen Mouse Brain Atlas ISH section data (section-to-CCF alignments)serial in situ hybridisation and Nissl sections (2D) aligned to the Allen Common Coordinate Framework reference volume (3D), between mice39,148 section datasets (API count), about 20,000 genes, sagittal and coronalper-section 2D and per-dataset 3D transforms from Allen's automated alignment pipeline, served by the API (a reconstruction, not expert landmarks)both imagesdirectAllen Institute Terms of Use (non-commercial)non-commercialwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
real; truth is the pipeline's alignment, known to be imperfect locally; overlaps the DeepSlice entry in purpose Licence checked on 2026-10-10 at https://alleninstitute.org/terms-of-use/; derived test pairs shareable: yes; commercial use: no. terms cover brain-map.org: noncommercial use, copying, distribution and derivatives allowed with citation; using the content to perform a commercial service for a third party needs written permissionexperts place 10 landmarks per section on 100 sections across genes, as DeepSlice's human-variability study did
AMBIA mouse section-to-atlas set (Sadeghi 2023)Coronal mouse brain sections (own fluorescence and Allen transgenic STPT) to Allen CCF/ARA; section-to-volume4,778 sections with expert atlas positions; subset with manual registrationThree experts' atlas section numbers; fully manual registrations by experts as ground truthboth imagesrequestdata on reasonable request; code MITaskneeds a request
the data needs a request or a data use agreement
Test data in the GitHub repo; full data on request. Licence checked on 2026-10-10 at https://pmc.ncbi.nlm.nih.gov/articles/PMC10406728/; derived test pairs shareable: unclear; commercial use: unclear. paper data statement: 'All data is available upon reasonable request'; repo licence MIT covers the code only; STPT images come from the Allen Institute
Apple slice photographs vs CTSlice photos vs CT of 1347 apples~1347 apples x several slicesRegistered CT slices (method output, not independent)directraw data CC BY 4.0 (Zenodo 8167285); results record 8275793 has no licenceopenwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Non-medical slice-to-volume Licence checked on 2026-10-10 at https://zenodo.org/records/8167285; derived test pairs shareable: yes; commercial use: yes. registered CT slices (truth) sit in record 8275793 with no licence field; rebuild pairs from the CC BY raw record or ask for the resultsTwo people click 10 points (core, seed tips, skin edge) on the photo and the matching CT slice of 20 apples; the registered slices are method output, not truth.
ASOCA (coronary CTA with centrelines, Coronary Atlas)40 annotated CCTA (20 normal, 20 diseased) with lumen masks, centrelines and meshes; used by DeepCA and NeCA to simulate two-view angiograms40 annotated + 20 test CCTA3-expert majority-vote lumen masks and manual centrelines; projected truth by constructionboth imagesrequestUK Data Service End User Licence; data files closed, released on depositor permissionaskneeds a request
the data needs a request or a data use agreement
Not a 2D/3D set by itself; earlier Zenodo records 4638863/4638864 list no licence status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Azad 2023 ViT review). Licence checked on 2026-10-10 at https://reshare.ukdataservice.ac.uk/855916/; derived test pairs shareable: no; commercial use: no. ReShare: ASOCA_Normal/Diseased.zip closed access, email asoca@unsw.edu.au then UKDS helpdesk; the EUL forbids sharing with anyone not registered with UKDS and commercial use without permissionSimulate two-view angiograms at clinical angles with recorded geometry for exact 2D/3D truth
Autoscoper training data (Queen's Skeletal Observation Lab)biplanar videoradiography to CT bone volumes, foot, knee and shoulder, same subjecta few trials per joint (count unverified)reference bone tracking transforms made in Autoscoperboth imagesdirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Tutorial package, small; real in vivo biplanar X-ray video.two people track each trial to give tracking spread
Berea sandstone 2D thin section in 3D CTOptical thin section vs binary micro-CT volume (plus synthetic sphere pack)Few pairsSynthetic case known; real case none stateddirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Non-medical slice-to-volumeThe synthetic sphere pack has known truth; for the real thin section, two people place 10 pore and grain points on the section and the matching CT slice.
Bi-plane EOS X-rays before and after total hip arthroplasty (UNIGE, Yareta)Biplanar standing EOS radiographs of pelvis and femur, same patient before and after hip replacement; no CT28 patients x 2 time points (biplanar)Anatomical landmark positions and 3D pelvis/femur parameters measured 3 times by 2 operators (sterEOS)both imagesdirectCC BY-NC-SA 4.0, plus a contract on the use of datanon-commercialwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Fails the 3D-volume criterion: no CT; usable as pre/post biplanar X-ray pairs with repeated landmarks; whether landmarks are 2D image coordinates or 3D only is unverifiedUse the repeated operator landmarks as pre/post truth; for 2D/3D, fit a statistical pelvis model and report landmark reprojection error
BIMCV-COVID19+ (chest X-ray and CT of the same patients)Real chest radiographs (CR, DX) and CT of the same COVID-19 patients over days to weeksIterations list up to 21,342 CR, 34,829 DX and 7,918 CT studies; patients with both modalities unverifiedNone for registration (radiological findings, 23 segmented images)one imagerequestBIMCV-COVID19 Dataset Research Use Agreementnon-commercialneeds a request
the data needs a request or a data use agreement
Weak pairs: different days, positions and disease state; NIfTI in MIDS format Licence checked on 2026-10-10 at https://bimcv.cipf.es/bimcv-projects/bimcv-covid19/bimcv-covid19-dataset-research-use-agreement-2/; derived test pairs shareable: no; commercial use: no. free for research only, sale prohibited, non-academic use approved case by case; no distribution without written permission and no derivative works (except models); colleagues in the same organisation may use it under the same agreementSelect patients with CXR and CT within 24 h; two people mark 10-15 bony and airway landmarks on both; score projected landmark error
Biplanar videoradiography foot dataset (Welte 2022)biplanar X-ray video (250 Hz) of a running and hopping foot vs CT of the same foot (tibia, talus, calcaneus), in vivo1 participant, several running and hopping trials; 41-frame assessment trialbone poses from implanted tantalum beads (3-5 per bone), beads then removed digitally from X-ray and CTboth imagesdirectnot found; article CC BY-NC-ND 3.0; code AGPL-3.0unknownwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Data link is in the JoVE article (not checked); code at github.com/skelobslab/JOVE_BVR_FootModelAndMarkerBased. Licence checked on 2026-10-10 at https://doi.org/10.3791/63535; derived test pairs shareable: unclear; commercial use: unclear. JoVE page blocks scripted reads, the companion code repo (skelobslab/JOVE_BVR_FootModelAndMarkerBased) gives no data link, and no DataCite/Zenodo record was found; ask Rainbow lab (Queen's)
Biplanar X-ray to 3D shape benchmark (DRRs from 6 public CT sets)AP and lateral DRRs generated from TotalSegmentator, CTPelvic1K, VerSe, RSNA cervical fracture and other CTs for femur, pelvis, vertebra and rib shape reconstructione.g. 465 femur, 446 + 1,106 pelvis, VerSe vertebrae, 710 cervical vertebraeDRRs and 3D bone masks aligned by construction; clinical landmarks extracted automaticallyboth imagesdirectcode GPL-3.0; no data distributed, source CT sets keep their own licencesaskwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Authors state no public paired real X-ray + CT exists; repository URL not opened Licence checked on 2026-10-10 at https://github.com/naamiinepal/xrayto3D-benchmark; derived test pairs shareable: unclear; commercial use: unclear. DRRs are regenerated with XrayTo3DPreprocess from the source CT sets (VerSe etc.), whose terms must be accepted one by one; pick open sources to make it usableUse the toolkit's DRR geometry with random rigid offsets for a biplanar 2D/3D pose task
Bite2Text (intraoral scans with paired intraoral photographs, ODIN 2026)intraoral colour photographs (2D) and upper/lower intraoral surface scans (3D mesh) of the same patients1,000 patients: 2,000 registered upper/lower scans, 5,000 paired photographs (paper); ODIN lists 2,000 training cases and 50 hidden test casesnone for photo-to-scan registration (scans registered to each other; reports by clinicians)one imageaccountCC BY-NC-SA (ODIN 2026 challenge page)non-commercialneeds an account
the download needs an account
real; photo-to-mesh pose is a clean rigid 2D/3D problem with exact geometry on the scan sidetwo people click 8 tooth landmarks (cusp tips, incisal edges) on photos and on the scan for 50 patients; PnP from the clicks gives a reference pose
Boston Children's neonatal Vein of Galen DSA-MRA set (xvr)Intraoperative 2D DSA to preoperative MRA in neonates; same patient4 neonatesNot specified in the excerpthiddenNot statedunknownleft out
its images or its truth are not released
PrivateTwo people mark 10 vessel bifurcations on DSA and MRA; poses by PnP
BREAST-DIAGNOSIS (TCIA, mammography with breast MRI and CT)screening/diagnostic mammograms (2D) and breast MRI or CT (3D) of the same patients88 patients; 14 with both MG and MR/CT (NBIA series query)none for registrationone imagedirectCC BY 3.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
real; small usable subset; compression differs between mammography and MRI (prone, uncompressed), so a deformable 2D/3D problem. Similar tiny pools: TCGA-BRCA (3 MG+MR patients), RIDER Pilot (5 CR/DX+CT), VAREPOP-APOLLO (7 CR/DX/RF+CT/MR)two radiologists mark the lesion centre, nipple and 3-5 calcifications or clips in both views; clip markers are the strongest truth
BREATH (in vivo bronchoscopy localization dataset)Routine clinical bronchoscopy videos with patient-specific airway meshes from preoperative CT66 proceduresCalibration, depth, 6-DoF camera poses and airway skeletons with anatomical labels (pose source unverified)both imageshiddenunknownunknownleft out
its images or its truth are not released
Release status not stated in the text seen; largest in vivo video-to-CT set if releasedIf released: two people verify 10 branch-point frames per case against the CT airway labels
Brigham neurosurgery DSA-CTA cohort (xvr)Intraoperative 2D DSA to preoperative CTA, neurosurgery; same patient122 DSAs, 50 CTAs (50 patients)Manual keypoints with PnP-derived C-arm posesboth imageshiddenNot statedunknownleft out
its images or its truth are not released
Private
BronchoPose (synthetic bronchoscopy navigation from CT airways)Simulated bronchoscopy trajectories rendered in CT airway models of several patientsunverifiedExact poses by construction (flat-shaded rendering)both imageshiddenunknownunknownleft out
its images or its truth are not released
Paper promised release on acceptance; download not foundpair with real bronchoscopy of a CT-scanned phantom to test domain transfer
BronchOpt synthetic bronchoscopy benchmark (AeroPath CTs)Synthetic bronchoscopy frames rendered inside airway meshes from 15 public AeroPath CTs, frame-to-CT pose recovery from a perturbed start15 sequences (bronch_real_aero1-15) at 3 difficulty levelsExact 7-DoF poses, depth and airway mesh per frame by constructionboth imagesdirectrepo Apache-2.0; dataset (JHU SharePoint) no separate licenceaskwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Static anatomy, rendered appearance; AeroPath CTs give the 3D side Licence checked on 2026-10-10 at https://github.com/Soooooda69/BronchOpt; derived test pairs shareable: unclear; commercial use: unclear. README License section points to the Apache-2.0 LICENSE without saying whether it covers the off-repo dataset; renders built from AeroPath CTsregenerate from AeroPath CTs with a renderer: exact poses
C-arm CBCT spherical fiducial phantom studies (Yaniv 2009)C-arm cone-beam projection cine loops and their 3D reconstructions, two phantoms with spherical fiducials2 phantom studies, each a full projection sweep plus reconstructionSystem projection matrices for every projection; fiducials visible in 2D and 3Dboth imagesdirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Small, DICOM, MATLAB readers included; ideal sanity test for 2D/3D pose with exact geometryMove the reconstruction by a known rigid transform and ask methods to recover it from 1-3 projections; fiducials give independent TRE
C3VDcolonoscopy video to CT model, phantom22 videospose and depth from 2D/3D registrationboth imagesdirectCC BY-NC-SA 4.0non-commercialwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
C3VDv2 (colonoscopy 3D video dataset v2)Colonoscopy video of 60 silicone colon phantom segments registered to their 3D models; also simulated screening runs and deformation videos192 sequences, 169,371 frames; 169 videos with full truth, 8 screening videos with poses6-DoF pose, depth, normals, flow, occlusion, coverage maps and 3D models from 2D/3D registration to the phantom modelsboth imagesdirectCC BY-NC-SA 4.0non-commercialwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
status change for C3VD: v2 is 8x larger with more realistic phantoms and screening trajectories
CAMS-Knee (instrumented knee implants with moving fluoroscope)moving single-plane fluoroscopy (2D) of total knee replacements to implant models (3D) during daily activities, same subject6 subjects, 5 activities, multiple cyclestibio-femoral implant kinematics from 2D/3D model registration (ETH), implant loads from instrumented prosthesisboth imagesrequestCAMS-Knee licence agreement signed by the institution; scientific use without commercial benefitnon-commercialneeds a request
the data needs a request or a data use agreement
real in vivo; truth is the provider's own 2D/3D fit, not independent Licence checked on 2026-10-10 at https://cams-knee.orthoload.com/data-request/; derived test pairs shareable: no; commercial use: no. terms: data only for use at the registered institute, may not be passed on to third parties; legal page forbids providing CAMS-Knee files from other databases; commercial use needs written permissionsecond registration by another group on a 100-frame subset to measure inter-method spread
CAVAREV (cardiac vasculature reconstruction evaluation)Simulated rotational C-arm angiography of the 4D XCAT phantom with contrasted coronaries, cardiac and breathing motion2 dynamic projection sets (periodic cardiac; cardiac + breathing)Real C-arm geometry calibration; phantom ground truth used by the online scorerboth imagesdirectnone stated; 'open (i.e. public) and easy accessible projection data'askwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Download site www.cavarev.com did not respond on 2026-10-10 (unverified access); phantom Licence checked on 2026-10-10 at https://www5.cs.fau.de/research/software/cavarev/; derived test pairs shareable: unclear; commercial use: unclear. FAU pages offer free downloads but name no licence or terms; ask the Pattern Recognition LabPhantom is known: per-frame 3D vessel model and geometry give exact 2D/3D truth
CBCT dataset of a knee phantom (Planmeca Viso G7)Clinical dental/ENT CBCT scanner projections of a custom knee phantom1 phantom, 500 projectionsPer-projection source and panel coordinates, angles, distances and panel tiltboth imagesdirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Raw (flat-field only) projections from a clinical scanner with full geometryExact by construction after a known rigid move of the reconstruction
CBCT projection images of phantoms (Birklein 2025)Three simulated polychromatic CBCT scans of numerical phantoms (one with scatter) and one real skull scan on a Morita 3D Accuitomo 1704 projection sets (3 synthetic, 1 real skull)Synthetic phantoms known exactly; real skull geometry from the scanner (format unverified)both imagesdirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Built for artefact correction, not registrationUse the numerical phantoms for exact 2D/3D truth; real skull needs a reconstruction plus a known move
Component-wise translated CT projections (Qure.ai ct2xr, CT-RATE)Synthetic chest radiographs from 21,887 CT-RATE volumes, decomposed into bone, soft tissue and lung projections, style-translated toward real CXR21,887 projections (one per CT)Projection of the CT by construction; per-component layers; geometry only as spacing fieldsboth imagesdirectCC BY-NC-SA 4.0non-commercialwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Projection geometry not stored beyond pixel spacing; translated images are no longer exact line integralsUse with CT-RATE volumes and re-render at recorded poses for exact truth; translated images test appearance robustness
Cone-beam CT dataset of a walnut (Helsinki 2022)In-house lab cone-beam projections of one walnut with geometry metadata1 object, 721 projections (0.5 degree steps)Scan geometry metadata fileboth imagesdirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Non-medical; companion to the FIPS 2D fan-beam setsReconstruct, apply a known rigid move, recover it from sparse projections
Cone-beam X-ray CT data collection of 42 walnuts (Der Sarkissian, CWI)Laboratory cone-beam projections of 42 walnuts on three source orbits, with reconstruction pipeline and high-quality reference volumes42 objects x 3 orbits (1,200+ projections each, unverified count)Exact acquisition geometry per projection; artefact-free reference reconstruction from the combined orbitsboth imagesdirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Non-medical but realistic natural variability; same object scanned on three orbits gives multi-view projection sets in one frameExact by construction: register a few projections of one orbit to the reference volume after a known rigid move; or walnut-to-walnut 3D labels
CTPelvic1K (source CT for pelvic DRR benchmarks)1,106 pelvic CTs with manual bone labels, used to render synthetic pelvic X-rays at known poses (XPos3R, biplanar shape benchmarks)1,106 CTs (any number of rendered views)Exact by construction once DRRs are rendered at recorded poses; 3D bone labels give projected masksboth imagesdirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Not a 2D/3D dataset by itself; best open source of volumes for synthetic pelvis 2D/3D with a commercial-friendly licence Used in 1 of the 2D and microscopy registration papers read (XPos3R 2026 (training)).Render DRRs with DiffDRR or DeepDRR at random C-arm poses and store projection matrices; test on real DeepFluoro X-rays for sim-to-real
CTSpine1K (source CT for spine DRR benchmarks)1,005 spine CTs with vertebra labels, a source of volumes for synthetic spine X-ray / fluoroscopy at known poses1,005 CTs, 11,000+ vertebrae (rendered views unlimited)Exact by construction once DRRs are rendered at recorded poses; vertebra labels give projected masks and per-level posesboth imagesdirectCC BY-NC-SA (inherited from sub-datasets)non-commercialwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Not a 2D/3D set by itself; no public real spine fluoroscopy with CT was found (RadGS-Reg, SGReg, DVAP-Reg use in-house real data)Render AP/lateral DRRs per vertebra level with DiffDRR at random poses; add per-vertebra rigid perturbations for multi-body 2D/3D truth
CTSpinoPelvic1K (spine, pelvis, ribs, femora in one frame)CTs with spine, pelvis, ribs and femora labelled in one coordinate frame (lumbosacral transitional anatomy), usable as source volumes for multi-bone DRRsabout 1,000 CTs (count unverified)Multi-bone labels; projected truth exact by constructionboth imagesdirectCC BY-NC-SA 4.0non-commercialwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Labels over existing public CTs (CTPelvic1K/CTSpine1K sources); images must be fetched from the original collectionsRender multi-bone DRRs with independent per-bone rigid moves (polyrigid 2D/3D truth, as in PolyPose)
CZ CryoET Data Portalcryo-electron tomography tilt series, projection-to-projection and projection-to-tomogram alignmentthousands of tilt series across many datasetsalignment files from processing; gold fiducials in many datasets; annotations on tomogramsone imagedirectmostly CC BY 4.0 / CC0 per dataset (unverified)openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Alignment parameters are method outputs; gold beads are physical markersgold bead positions tracked across tilts as fiducial truth
DeepFluoroCT to fluoroscopy, pelvis of 6 cadavers, same specimen366 X-raysposes from offline registration, 14 CT landmarks, 2D segmentationsboth imagesdirectCC BY-NC 4.0non-commercialwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Unberath 2021 2D/3D review). Used in 4 of the 2D and microscopy registration papers read (DiffPose 2024; xvr 2025; PolyPose 2025; XPos3R 2026).
DeepSlice ground-truth alignmentsMouse coronal histology sections to Allen CCF7 section seriesHuman QuickNII alignments (affine-like to atlas)directCC BYopenwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
A second person redoes QuickNII on 20 sections to measure inter-observer error; the alignments are affine to the atlas, not deformable truth.
DFIS tibiofemoral kinematics (Mendeley)dual fluoroscopy to MRI bone models, knee during loaded walking and running26 adults, 4 conditionstracked bone kinematics (whether images are included: unverified)directCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
May hold kinematics only.Check that the images are included; if so, the tracked bone poses give 2D/3D truth, checked by a second person on 20 frames.
DRR-RATESynthetic AP and lateral chest radiographs (DRRs) made from CT-RATE chest CT, same patient by construction50,188 CT volumes (25,692 scans, 21,304 patients), each with AP and lateral DRRExact by construction: Siddon-Jacobs ray-traced DRR at documented default geometry (300 mm shift, lateral = 90 degree rotation)both imagesdirectCC BY-NC-SA 4.0non-commercialwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
PNG images with no stored projection matrix; geometry must be rebuilt from the documented parameters; DRRs, not real X-raysRe-render DRRs with DiffDRR or DeepDRR at random recorded poses from the CT-RATE volumes for exact 6-DoF truth
Dual-modal CT/US kidney phantoms (Gomes-Fonseca 2017)ultrasound (stored as .mha volumes, likely tracked sweeps; unverified) and CT of 10 kidney phantoms for percutaneous renal access10 phantomsfiducial markers localised in both CT and US, with surfaces, segmentations and aligned 3D Slicer scenesboth imagesdirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
phantom; may be 3D/3D rather than 2D/3D if US is a reconstructed volume
ETH moving fluoroscope knee implant images (Taylor lab)Single-plane moving fluoroscope images of TKA implants during daily activities, implant CAD modelsabout 106,000 annotated imagesAnnotated 6D implant poses from the lab's registration pipelineboth imagesrequestconfidential; access on request under Swiss law, ethics and NDAsaskneeds a request
the data needs a request or a data use agreement
Public phantom counterpart is Veriserum Licence checked on 2026-10-10 at https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0270596; derived test pairs shareable: no; commercial use: unclear. paper data statement: x-ray and implant data confidential, request via bt@ethz.ch under NDAs; only trained models are in the ETH repository (doi 10.3929/ethz-b-000544626)
FaBiAN fetal brain MRI simulator + synthetic datasetssimulated 2D T2w fast spin echo fetal brain slices with motion (2D) from a high-resolution segmented fetal brain model (3D)unlimited by simulation; published sets across gestational ages (Lajous et al.)known motion and source volume by construction (numerical simulator)both imagesdirectBSD-3-Clause (code)openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
synthetic, more realistic MR physics than slice extraction; generator rather than fixed datafix a seed set of 20 simulated subjects with logged slice poses as a frozen benchmark
FACE-R (head CT and upright 3D face scans, Hungary)upright 3D facial surface scans (3D surface, photograph-like) and supine head CT (3D) of the same people400 clinical patientsnone released; authors aligned face scans to CT skin surfacesone imagerequestnone stated; research database not publicaskneeds a request
the data needs a request or a data use agreement
real; would test surface-to-CT registration with gravity deformation; no data route found. No public paired face-scan/CT set was found in this sweep Licence checked on 2026-10-10 at https://real.mtak.hu/12324/; derived test pairs shareable: unclear; commercial use: unclear. OTKA project report describes the FACE-R MySQL database (425 living subjects, 5 cadaver heads) with no access route or licenceif obtained: 5 facial landmarks (endocanthion, pronasale, cheilion) placed by two people on scan and CT skin
FDA AO-OCT photoreceptor and RPE macular topography dataset (RST26OM03.01)AO-OCT volumes of the outer retina at 9 locations, 11 healthy eyes; 7 re-imaged on two more days at 2 locations7 subjects x 2 locations x 3 sessions repeat pairs; 10 volumes per videoPR and RPE cell markings by two graders (per session, not tracked across sessions)one imagedirectFDA regulatory science tool, open access via Synapse syn63546986 (Synapse login likely)openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
authors made no attempt to track cells across sessions; volume-to-volume within a video is a speckle-averaging registration taskmatch RPE and cone markings across sessions after rigid fit, person verifies
FDTooth (intraoral photographs with CBCT)intraoral colour photographs (2D) and CBCT (3D) of the same patients' anterior teeth241 patients, 1,800 annotated tooth boxesfenestration/dehiscence labels on photos derived from CBCT; no photo-to-CBCT registrationone imagecontrolledPhysioNet credentialed access (DUA)askneeds a request
the data needs a request or a data use agreement
real; photo-to-CBCT surface registration is possible via tooth crowns segmented from CBCTsegment crowns in CBCT, then two people click 6 crown landmarks per tooth on photos and CBCT surface for 30 patients
Femur multi-view X-ray/CT (Balgrist, bone substructure contours)CT to real intraoperative X-rays of the femur, 5 cadaveric specimens, same specimen5 specimens, 49 X-rays (count from the xvr-data copy)X-ray poses from a custom calibration fiducial marker; bone substructure contour masksboth imagesdirectCC BY-NC-ND 4.0 (xvr-data copy)non-commercialwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Original download location not seen; remixed DICOM/NIfTI copy hosted with permission in huggingface eigenvivek/xvr-data Used in 1 of the 2D and microscopy registration papers read (xvr 2025).Pose truth exists; add 10 bone landmarks marked on CT and on two views by two people to get an independent reprojection check
FIPS open X-ray tomographic datasets (walnut, lotus root, carved cheese, emoji)Small open X-ray tomography sets from the Finnish Inverse Problems Society, mostly 2D fan-beam sinograms4 objects (records 1254206, 1254204, 1254210, 1183532)Known fan-beam geometryboth imagesdirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Mostly fails the 3D criterion (2D fan-beam); non-medicalToy 1D-to-2D registration by construction only
FluoroSeg (FluoroSAM synthetic X-ray set)Synthetic X-ray and fluoroscopy images simulated from 1,621 whole-body CTs (New Mexico Decedent Image Database) with surgical tools, same subject by constructionabout 2.95 million images from 1,621 CTsProjected masks of 128 TotalSegmentator organs and 464 tools; C-arm geometry sampled per image (whether poses ship is unverified)both imagesaccountCC BY-NC-SA 4.0non-commercialneeds an account
the download needs an account
Gated on Hugging Face (accept terms); card unreadable without login; CT volumes come from NMDIDIf poses are in the scene files, exact 2D/3D truth by construction; otherwise re-simulate with DeepDRR from NMDID CTs
GAMMA (glaucoma fundus photo and OCT volume pairs)colour fundus photograph (2D) and macular OCT volume (3D) of the same eye300 pairs in the challenge (public release possibly 100; sources disagree)none for registration; fovea coordinates and disc/cup masks on fundusone imageaccountCC BY-NC-ND (per Hugging Face mirrors; grand-challenge terms not read)non-commercialneeds an account
the download needs an account
real; fails the truth criterion; NoDerivs may forbid sharing resampled imagesvessel bifurcations on the OCT en-face projection and on the fundus, found by a program and verified by a person, 10 per pair on 30 pairs
GIMM synthetic multi-view coronary angiography (from ImageCAS)Synthetic coronary angiograms (RCA and LCA) projected orthographically from ImageCAS CCTA under clinical C-arm angles, view pairs of the same patient26,205 multi-view image pairs, patient-split train and held-out testExact by construction: per-pair view angles, intrinsics K1/K2, 3D centreline points and their 2D projections; 20 matched landmarks per viewboth imagesdirectcode Apache-2.0; DRRs derived from ImageCAS (CC BY-NC 4.0)non-commercialwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Orthographic, not perspective, projection; images are vessel-mask DRRs, not real angiograms; regeneration code included Licence checked on 2026-10-10 at https://github.com/medipixel/GIMM; derived test pairs shareable: yes; commercial use: no. README warns the DRRs inherit ImageCAS terms; ImageCAS on Kaggle is CC BY-NC 4.0, so the derived DRRs are non-commercialRegenerate with perspective C-arm geometry and full-CT DRRs from ImageCAS to get a 2D/3D CCTA-to-angiogram task with exact pose truth
Henninger shoulder biplane fluoroscopy kinematics (Utah, Zenodo)Biplane fluoroscopy (100 Hz) of humerus and scapula tracked with CT bone models, healthy and post-surgical shouldersDozens of subjects and trials (20 healthy in the first release; rTSA set adds 10 + 10)Bone poses from semi-automatic DRR-to-frame matching (model-based markerless tracking) by the labone imagedirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Fails: archive sizes suggest kinematics, landmarks and transforms only, no radiograph frames or CT (unverified)Ask the lab for raw biplane frames and CT for a few trials; the published poses then become reference poses
Hip joint 2D-3D gold standard (HES-SO Geneva)CT of a female hip phantom to fluoroscopy views1 CT, 19 fluoroscopy images at different angles and voltagesfiducial-based transforms with uncertainty propagation (uTRE)both imagesdirectCC BY-SA 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Phantom, one object; large rotation angles. Licence checked on 2026-10-10 at https://doi.org/10.26037/yareta:2kw5s2jadbbv3ngtkzsovihtnu; derived test pairs shareable: yes; commercial use: yes. same Yareta archive (ed4bb43a) as the hip phantom row; Yareta metadata says public, data tag BLUE, aip-dua false (no data use agreement); share-alike on derived pairs
Hip phantom 2D-3D registration gold standard (D'Isidoro, Schmid; Yareta)CT to fluoroscopy, female hip phantom, large rotation angles, same object1 CT, 19 fluoroscopic views at different angles and voltagesFiducial-based transformations jointly optimised with noisy 3D fiducials (MPPC) plus an uncertainty-based TRE (uTRE), by the authorsboth imagesdirectCC BY-SA 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Phantom; fiducials visible in images (mask or inpaint before scoring); companion paper gives a standardised protocol Licence checked on 2026-10-10 at https://access.yareta.unige.ch/access/metadata/ed4bb43a-eaaa-4e62-8763-bd5803f5cd47; derived test pairs shareable: yes; commercial use: yes. Yareta metadata: access PUBLIC, data tag BLUE, aip-dua false, so the earlier 'contract on the use of data' is not on this archive; paper (Med Phys 2021) itself is CC BY-NC; share-alike on derived pairs
HKU CBCT projection dataset (hku-mip/cbct_projection)CBCT projections of brain, head and neck, lung, prostate and pancreas cases in three system geometries (Elekta XVI, Varian TrueBeam, IBA/Toshiba)100,000+ files, 720 projections per case; cases named after TCIA collections (HN-CHUM, Pancreas-CT-CB, Prostate-AEC)System geometry per scanner (reader and RTK FDK scripts in github.com/MedIP-HKU/CBCT_projection)both imagesdirectMITopenwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Card says acquired on 3 systems, but identical TCIA case names across systems suggest simulation from planning CT (unverified); if simulated, the source CT gives exact truthIf simulated from TCIA CT: exact 2D/3D truth by re-projecting the CT; otherwise register projections to the TCIA planning CT and check with bony landmarks
ImageCAS (coronary CTA source for synthetic angiography)1,000 coronary CTA scans with left and right coronary masks; source volumes for projected angiograms (GIMM, DeepCA, NeCA)about 1,000 CCTA (rendered views unlimited)3D coronary masks; projected truth exact by constructionboth imagesrequestCC BY-NC 4.0non-commercialneeds a request
the data needs a request or a data use agreement
Not a 2D/3D set by itself; centreline add-ons: ImageCAS-X (Zenodo 21887809, CC BY 4.0, 800 scans) and Awakening ImageCAS (Zenodo 23243094, CC BY-NC 4.0, ostia + 999 centreline trees) Licence checked on 2026-10-10 at https://www.kaggle.com/datasets/xiaoweixumedicalai/imagecas; derived test pairs shareable: yes; commercial use: no. licence of the authors' own Kaggle upload (Kaggle API licenseName); GitHub repo has no licence and gives the data by emailProject CCTA with a perspective C-arm model at recorded angles; use the published centrelines for projected-centreline error on real angiograms if any are paired later
IOERT pig limb C-arm projections and CT (IEEE DataPort)C-arm projection radiographs (2D) to CT and CBCT (3D) of a pig lower limb with an electron radiotherapy applicator1 phantom, 7 projections per condition (with and without soft tissue)fiducials on applicator and skin; applicator pose matrices providedboth imagesaccountCC BY 4.0openneeds an account
the download needs an account
real phantom; small; free IEEE account needed Licence checked on 2026-10-10 at https://doi.org/10.21227/gxpd-d998; derived test pairs shareable: yes; commercial use: yes. DataCite rights for the IEEE DataPort open-access record; download needs a free IEEE account
JHU OCT multiple sclerosis and healthy controls with manual boundariesSpectralis macular OCT volumes, 14 controls and 21 MS patients35 volumes, one per subjectmanual delineation of nine retinal boundariesdirectunverifiedunknownwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
fails the pair criterion (single visit); useful only for between-subject or synthetic-warp testsknown smooth deformation applied to each volume, boundaries give exact truth
KidneyDepth (synthetic and real ureteroscopy)rendered ureteroscopic frames (2D) inside kidney collecting-system meshes from CT (3D), plus real ureteroscopy frames without truthsynthetic sequences for depth and SLAM trajectories; real frames for visual checkdepth and camera trajectory known by construction (BlenderProc rendering of CT meshes)both imagesdirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
synthetic truth only; record text still says 'full dataset will be released upon acceptance' although marked openCT-scanned silicone kidney phantom with an EM-tracked ureteroscope (as in RAUL) for real frames with pose
Knee Grand Challenge (Grand Challenge Competition to Predict In Vivo Knee Loads)single-plane fluoroscopy of an instrumented knee implant during gait (2D) to implant CAD and CT bone models (3D), same subject6 subjects with instrumented TKA; treadmill gait fluoroscopy trialsimplant and bone kinematics from 2D/3D model-image registration by the data providers (Banks); implant loads measured by the instrumented prosthesisboth imagesdirectMIT Use Agreement (SimTK packages)openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
real; registration truth is itself a 2D/3D fit (no independent markers); implant geometry exact, which makes it a good rigid 2D/3D test Licence checked on 2026-10-10 at https://simtk.org/frs/?group_id=413; derived test pairs shareable: yes; commercial use: yes. each SimTK download package (CT, X-ray, motion, geometry) shows 'View License' = MIT Use Agreement; checked for the 4th-6th competition packagescheck provider kinematics with a second registration tool and report the spread; implant CAD gives exact silhouettes
Knee radiograph to CT in-house cohort (dense 2D-3D correspondence, 2026)Plain knee radiographs (AP, lateral) to the same patient's segmented CT; femur and patella pose49 radiographs from 26 patients; plus 766 CTs for DRR trainingPer-view femur poses from an in-house tool, operator-aligned and annotator-validatedboth imageshiddenNot statedunknownleft out
its images or its truth are not released
Private
Laparoscopic liver registration ground truth by LUS (Rabbani et al., EnCoV)preoperative 3D liver and tumour model to laparoscopic images, with tracked laparoscopic ultrasound as ground truth, same patient4 patients, 4 tumourscalibrated LUS tumour profiles in laparoscope coordinates, landmarksboth imagesdirectGNU GPL v3 for research purposes onlynon-commercialwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Tiny but clinical.
LDCT-and-Projection-data (TCIA, DICOM-CT-PD)Raw helical CT projection data with acquisition geometry plus reconstructed CT, head, chest and abdomen, full and simulated low dose299 subjects (199 patients + phantom public; 99 head restricted)Vendor-decoded acquisition geometry per projection in DICOM-CT-PD; reconstructions in the same frameboth imagesdirectCC BY 4.0 (chest, abdomen, phantom); head under NIH controlled accessopenwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Helical fan/cone geometry, not C-arm; Helix2Fan rebins to flat detector status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Foundation models review 2026).Exact by construction: project-to-volume pose recovery on subsets of projections, or register low-dose projections to a moved full-dose volume
LIDC-IDRI (CT + chest radiograph pairs)Real chest radiographs and thoracic CT of the same patients, different sessions and positions290 chest radiographs among 1,010 patientsNone for registration (nodule marks on CT only)one imagedirectCC BY 3.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Weak pair: upright PA radiograph vs supine CT; acquisition geometry not recorded; NIfTI mirrors (MedOtter/LIDC-IDRI) skip the radiographs status change: Round 4 count: listed in 4 of the 38 round-4 surveys and benchmark papers (Boveiri 2020 review; Bharati 2022 review; Chen X 2021 review (Prog Biomed Eng); Foundation models review 2026).Two people mark 10-15 landmarks (clavicle heads, carina, rib crossings, nodules) on radiograph and CT; score reprojection error after rigid plus breathing deformation
LIDC-IDRI-DRRDRRs and projected bone/rib label maps made from LIDC-IDRI chest CT, same patient by constructionunverified count (derived from LIDC-IDRI CTs)Projected rib segmentations aligned with each DRR by constructionboth imagesdirectCC BY 3.0 Unportedopenwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Built for segmentation; projection geometry stored or not is unverified; download link not opened Licence checked on 2026-10-10 at https://sites.google.com/view/virginiafernandes/datasets/lidc-idri-drr; derived test pairs shareable: yes; commercial use: yes. page: 'Users of this data must abide by the Creative Commons Attribution 3.0 Unported License'; also asks users to contact oliveirahugo@dcc.ufmg.br and cite the paperRe-project the CT with recorded random poses to get exact 2D/3D truth
Ljubljana 2D-3D-GSCT and MR to 18 X-rays, cadaver lumbar spine1 specimenfiducial gold standardboth imagesdirectCC BY-NC-NDnon-commercialwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Ljubljana 3D-2D-GS-CA3D-DSA to 2D-DSA, cerebral vessels, same patient10fiducial gold standardboth imagesdirectCC BY-NC-NDnon-commercialwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Used in 2 of the 2D and microscopy registration papers read (DiffPose 2024; xvr 2025).
MATCH (Markerless lung target tracking AAPM Grand Challenge)intra-treatment kV and MV projections (2D) of a moving 3D-printed thorax phantom against its planning CT (3D)Part A in silico/phantom: 4 blinded patient-derived motion traces with VMAT plans; Part B experimental (scope not read)known phantom displacement from a HexaMotion stage driven by patient lung tumour tracesboth imagesdirectnone stated; 'Anyone is welcome to use the MATCH Challenge datasets for their research', cite the referenceaskwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
phantom with exact motion truth; Part A on the University of Sydney repository (handle 2123/34376) Licence checked on 2026-10-10 at https://aapm.org/GrandChallenge/MATCH/; derived test pairs shareable: unclear; commercial use: unclear. data on Google Drive, analysis code on GitHub with no licence; research-use wording only
Moorfields AMD Dataset 002 (INSIGHT, longitudinal OCT and colour photos)Topcon and Heidelberg macular OCT volumes, colour and other photos over anti-VEGF treatment visits, same eye over time19,785 eyes, 2.9 million images, many visits per eyenone for registration; injection and visual acuity recordscontrolledHDR UK gateway, data access request via INSIGHT hubaskneeds a request
the data needs a request or a data use agreement
Dryad releases (Moraes 2020, 2024) hold only AI-derived volumes per visit, not imagesvessel shadows on OCT en-face projections give bifurcation landmarks between visits, found by a program and verified by a person
MR CLEAN Registry CTA and DSA (stroke thrombectomy)Cerebral CTA to 2D DSA (AP and lateral), pre and post thrombectomy, same patient182 patients with AP and lateral DSA in one study; 94 lateral pairs usedManual rigid transformation matrices by an in-house researcher (Erasmus MC); DSA pre/post point correspondences in a second studyboth imagescontrolledRegistry data, not publicaskneeds a request
the data needs a request or a data use agreement
Fails access: listed because it is the largest known CTA-DSA set with manual pose truthIf access is granted: two people align 20 test pairs to measure inter-observer pose error
Multimodal 3D point clouds and CT volumes (Saiti 2022)Real micro-CT volumes and synthetic point clouds generated from them under known transforms, for point-cloud-to-volume alignmentunverified count (training arrays only)Ground-truth transformation matrices, exact by constructionboth imagesdirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Surface-to-volume, not image-to-volume; subject of the micro-CT scans unverified (likely non-medical objects)
Multimodal dental dataset (PhysioNet; CBCT, panoramic, periapical)Dental CBCT with same-patient real panoramic radiographs and periapical radiographs simulated from CBCT (Siddon ray tracing)v1.1: 329 CBCT from 169 patients, 8 panoramic radiographs (v1.0: 13), 16,203 simulated periapical images from 188 CBCTSimulated periapical images aligned to CBCT by construction (3 angles per tooth); panoramic pairs have no truthone imagecontrolledPhysioNet Contributor Review Health Data License 1.5.0 (credentialed, DUA, study review)askneeds a request
the data needs a request or a data use agreement
Only 8 real panoramic-to-CBCT pairs, acquired on a later visit; panoramic geometry (rotational slot) is not a simple projectionTwo people mark 10 dental landmarks (cusps, apices, implant tips) on panoramic and CBCT; or re-simulate panoramic views from CBCT with a known trajectory
Museum X-ray radiographs of objects with metal markers (Bossema, CWI)Raw radiographs of a wooden block (and a bronze sculpture) taken with basic museum X-ray equipment at several sites, plus a reference scan on the FleX-ray lab CTSame object at 4 facilities (Rijksmuseum, British Museum, Getty, FleX-ray) + Getty sculptureProjection geometry estimated from metal markers; FleX-ray reference scan has known scanner geometryboth imagesdirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Non-medical (cultural heritage); marker-based geometry calibration from 2D projections; sister records 8379910, 8379880, 10557034, 8379913Hold out some markers: compare geometry from the remaining markers against the held-out marker reprojection
Navigated EBUS bronchoscopy with CT (SINTEF, Sorger 2018)Electromagnetically tracked endobronchial ultrasound (convex probe) images fused with preoperative CT, lung cancer staging, same patient4 patients, tracked EBUS sequencesEM-tracked probe poses; target lymph nodes defined on CT and located in US (11 accuracy measurements, 10.0 +/- 3.8 mm)both imagesdirectCC0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Tracking error ~10 mm limits truth to a starting pose; contents per patient archive unverifiedTwo people mark lymph-node centres and vessel landmarks in US frames and CT; use the EM pose as initialisation only
Needles & Haystacks (domain-agnostic rigid slice-to-volume benchmark)2D slices (2D) to be located in 3D volumes (3D) from many sources: medical CT/MR, materials micro-CT and others; rigid 6-DoF158,436 registration problems; 8,461 hidden test problems on an online benchmarkexact slice poses by construction (slices resampled from the volumes)both imagesdirectCC BY 4.0 (Zenodo); code Apache-2.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
synthetic slices from real volumes; mixed domains (medical and non-medical); test truth hidden behind the serveradd real cut sections (thin sections or histology) of a few of the scanned objects with fiducials to test beyond resampled slices
New Mexico Decedent Image Database (NMDID: whole-body CT with scout radiographs)Whole-body post-mortem CT of 15,000+ decedents with CT scout (localizer) radiographs; source volumes for FluoroSeg and Pelphix simulations15,000+ decedents, each CT with scout imagesScout radiographs are real projections acquired in the CT session and frame (scanner geometry in DICOM, unverified per case); DRRs exact by constructionboth imagesrequestNMDID Database Terms of Use (data use agreement): research or education onlynon-commercialneeds a request
the data needs a request or a data use agreement
Scout-to-CT is a real 2D/3D pair with near-known geometry; same idea applies to localizer series kept in TCIA/IDC CT collections Licence checked on 2026-10-10 at https://nmdid.unm.edu/resources/data-use; derived test pairs shareable: no; commercial use: no. terms forbid commercial use and any duplication, distribution, derivative works or making data available to third parties; one-year access, destroy or return on request; account and approved request neededScout-to-CT: reconstruct the scanner slot geometry from DICOM tags and check with 10 bony landmarks marked on scout and CT
Numerical simulation of 3D/2D coronary artery registration (figshare, Chen 2025)Simulated data for 3D (CTA) to 2D (X-ray angiography) coronary artery registrationunverified (record has no description)unverified (simulation implies known transforms)directCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Record description is a placeholder; contents not inspected (no large downloads)If centrelines and projection parameters are inside, exact truth by construction; otherwise regenerate from ImageCAS/ASOCA
OpenNeuro ds004332 (brain MRI with and without deliberate head motion, markerless tracking)clinical brain MRI protocol (2D and 3D sequences) acquired still and with instructed nodding, same subject; slice or volume pose against a still reference22 healthy adults, 6 sequences, 4 scans each (2 still, 2 with motion)head pose time series from a markerless optical tracker (Tracoline, TracInnovations) shipped as motion transformsboth imagesdirectCC0 (raw k-space on PublicnEUro under a DUA)openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
real motion with an independent tracker: the only open brain set where 2D slices can be scored against measured pose; tracker-to-scanner cross-calibration error is part of the truthregister each 2D slice of the moving scans to the still 3D scan and compare with tracker poses; add 5 skin-marker or fat-navigator checks
P2I-LReg (in vivo laparoscopic liver preoperative-to-intraoperative registration)laparoscopic liver resection video keyframes (2D) to preoperative CT liver models (3D), same patient346 keyframes from 21 patients (15 new, 6 reprocessed from P2ILF)liver masks by six surgeons (4 label, 2 verify), reconstructed intraoperative point clouds, camera intrinsics; no independent pose truthone imagehiddennot stated; paper says code and dataset available via the Self-P2IR project (link not found)unknownleft out
its images or its truth are not released
real; evaluation is Dice of projected model against masks, a weak proxysurgeons mark ridge and ligament landmarks in frames and on the CT model (as P2ILF), or intraoperative ultrasound of tumours as EnCoV did
P2ILF (Preoperative to Intraoperative Laparoscopy Fusion, MICCAI 2022)preoperative 3D liver model (CT/MRI) to intraoperative laparoscopic image, same patient167 images from 9 patients (train), 16 images from 2 patients (test)2D and 3D anatomical curves (ridges, falciform ligament, silhouettes)both imagesrequestnone stated; patient training data released on request to challenge participants onlyaskneeds a request
the data needs a request or a data use agreement
Page says no longer accepting data requests; ask organisers. Licence checked on 2026-10-10 at https://p2ilf.grand-challenge.org/; derived test pairs shareable: unclear; commercial use: unclear. page: 'NO LONGER ACCEPTING ANY DATA REQUEST' since July 2022; contact the organisers
Pelphix simulated and cadaver pelvic fluoroscopy sequencesSimulated X-ray sequences of percutaneous pelvic fixation rendered (DeepDRR) from NMDID CTs, plus one real cadaver sequence279,709 simulated training + 8,515 validation images; 256 cadaver imagesSimulated: corridors, tools, anatomy and C-arm geometry by construction (pose files unverified); cadaver: phase labelsboth imagesdirectrepo code licence only (GitHub API MIT, README badge Apache-2.0); data (JHU SharePoint) no licence statedaskwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Built for phase recognition; NMDID source CTs needed for 2D/3D use Licence checked on 2026-10-10 at https://github.com/benjamindkilleen/pelphix; derived test pairs shareable: unclear; commercial use: unclear. simulated sequences are rendered from NMDID CTs, whose terms forbid derivative works, redistribution and commercial use; ask the authorsIf poses are stored: exact 2D/3D truth; otherwise re-simulate with DeepDRR at recorded poses
PENGWIN Task 2 (synthetic pelvic X-ray from CT)DeepDRR X-rays simulated from pelvic fracture CT (CTs in PENGWIN Task 1)100 CTs x 500 views = 50,000 images2D fragment masks; C-arm geometry sampled per image (whether projection matrices ship is unverified); 3D fragment labels in Task 1both imagesdirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Usable as 2D/3D with known pose only if geometry files are included
Pig lower limb CT to C-arm projections (IOERT, IEEE DataPort)CT to C-arm projection radiographs, pig lower-limb phantom1 CT, 7 projectionsskin and applicator fiducials used for point-based registrationboth imagesaccountCC BY 4.0openneeds an account
the download needs an account
IEEE DataPort needs a free account.
ProST cadaver X-rays (Gao et al. 2020)CT to real X-rays of a cadaver pelvis, same specimen10 real X-rays, 1 cadaverPoses from 1 mm metal beads injected into the bone surface, annotated on X-ray and CT, PnP solveboth imageshiddenunknownunknownleft out
its images or its truth are not released
Repo ships only an example cropped CT; real X-rays not public (a TIB index entry lists them with DOI 10.57702/4bmz85rj, download not seen)Ask the authors; bead truth is strong if released (beads must be masked before scoring)
R2-Gaussian CBCT datasets (synthetic and real)Cone-beam projections in NeRF format: synthetic (from public CT volumes, 512x512 projections, 256^3 volumes) and real (FIPS lab scans), with geometryabout 15 synthetic scenes x several view counts + a few real scans (unverified counts)Exact geometry in meta_data.json; source volumes for synthetic casesboth imagesdirectproject under the Gaussian-Splatting licence (non-commercial research and evaluation)non-commercialwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Reconstruction benchmark; also distributes SAX-NeRF scenes Licence checked on 2026-10-10 at https://github.com/Ruyi-Zha/r2_gaussian; derived test pairs shareable: unclear; commercial use: no. README: 'This project is under the license of Gaussian Splatting'; the data (Google Drive) has no separate licence; real scans derive from FIPS CT data and synthetic ones from public CT volumes with their own termsKnown rigid move of the volume, recover from sparse projections
RabbitCT (C-arm CT projections of a rabbit)Clinical C-arm cone-beam projections of a rabbit with calibrated 3x4 projection matrices, reconstruction benchmark496 projections (1248 x 960), 1 subjectCalibrated 3x4 projection matrix per projection; 256^3 reference reconstructionboth imagesdirectCC BY 4.0 (2026 archival re-release with the authors' agreement)openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Original FAU site offline; Zenodo copy has 512 and 1024 targets, reference volume and developer kit; preclinical, post-mortem rabbitReconstruct, perturb the volume pose, and ask methods to recover it from a few projections: exact 2D/3D truth by construction
RadGS-Reg in-house intraoperative spine biplanar setPreoperative spine CT to intraoperative AP/lateral X-rays; same patient10 cases, 30 vertebral-level samplesCT-to-C-arm pose recorded during C-arm circular acquisitionsboth imageshiddenNot statedunknownleft out
its images or its truth are not released
Private
Rat brain thionine/myelin serial sections (EBRAINS, Leergaard 2018)brain-wide serial coronal and sagittal rat brain sections (2D) for registration to the Waxholm Space rat atlas (3D)serial section sets (count not read)QuickNII/VisuAlign atlas anchoring by the EBRAINS tool authors used as example registrations (whether alignments ship is unverified)directCC BY-SA 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
real; tutorial dataset for 2D-to-3D atlas registration Licence checked on 2026-10-10 at https://api.datacite.org/dois/10.25493/C63A-FEY; derived test pairs shareable: yes; commercial use: yes. DataCite rights for the EBRAINS dataset 'Atlas of normal rat brain cyto- and myeloarchitecture'; share-alike; EBRAINS asks for citationtwo people anchor 10 sections each in QuickNII; inter-rater difference gives the error bar
Rat jaw micro-CT + histology (Rodgers)Synchrotron micro-CT vs histology slide, same jaw1 slice-to-volume pairNone stateddirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
One pair: two people click 10 points (tooth roots, canals) on the slide and the matching CT plane; a demo case only.
RealSynCol (synthetic colon from 10 CT scans)Realistic rendered colonoscopy frames inside colon geometries extracted from 10 CT scans28,130 frames over 10 colonsExact depth maps and camera poses by construction (Blender, with OpenCV conversion scripts)both imagesdirectnone stated for the data; paper on arXivaskwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Whether the source CT volumes ship is unverified Licence checked on 2026-10-10 at https://arxiv.org/abs/2602.08397; derived test pairs shareable: unclear; commercial use: unclear. data availability says 'The dataset is available at the following link: RealSynCol Dataset link' with no URL in the PDF; GitHub repo (code only) has no licence
RTI-3D auditable registration datasetReflectance transformation images vs projected 3D geometry of a carved relief46 RTI imagesFitting and 4 fixed validation landmarks, repeatability sessionsboth imagesdirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Cultural heritage; validation landmarks reused, not independent
SimCol3D (MICCAI 2022 EndoVis colonoscopy)Virtual colonoscopy rendered (Unity) inside three CT-derived human colon models, plus real EndoMapper patient sequences for relative poseSynthetic colons I-III with many trajectories (frame counts unverified); 7 real trajectories from 3 patients in Task 3Exact depth and camera pose per synthetic frame; colon I CT is public via EndoSLAM; test truth released after the challengeboth imagesdirectCC BY-NC-SA 4.0non-commercialwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Synthetic frames are registered to CT by construction; colons II and III CTs are from UCLH (not public, unverified) Licence checked on 2026-10-10 at https://doi.org/10.5522/04/24077763; derived test pairs shareable: yes; commercial use: no. UCL Research Data Repository record (DataCite rights); share-alikeadd a CT-scanned colon phantom with EM tracking to get real frames with exact pose
SVRTK fetal MRI datasets (simulated motion-corrupted fetal brain stacks)2D single-shot TSE slice stacks (2D slices) to a 3D fetal brain atlas volume; slice-to-volume reconstruction1 phantom case: 6 motion-corrupted stacks plus maskstacks simulated from a T2w fetal brain atlas, so the source volume and motion are known by construction (motion parameters shipped: unverified)both imagesdirectCC0 1.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
synthetic; one case only; atlas volume is public (CRL atlas, listed separately)simulate many more stacks from FeTA or dHCP volumes with recorded rigid slice motion (as SVoRT did) to get exact slice poses
Synthetic paired PA/lateral chest radiographs from LIDC-IDRI CTPA and lateral radiograph-like projections rendered from the same CT, with projected nodule masks683 PA/LAT pairsnodule masks projected from the same 3D segmentation (geometry from the renderer)both imagesdirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Synthetic; projection geometry files not checked.keep the rendering geometry to give exact 2D/3D pose truth
SyntheX hip imaging data (real and simulated X-rays)Real cadaver hip X-rays (DeepFluoro and extra Brainlab Loop-X images) and matched simulated X-rays from the same CTs, with segmentation and landmark labelsunverified counts (HDF5 real.h5, real_loopx.h5, realistic.h5 etc.)Segmentation and 14 anatomical landmarks per image; simulated images aligned to CT by constructionboth imagesdirectnot stated in the paper; data in JHU Research Data Repository (doi 10.7281/T1/2PGJQU); code MITaskwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Data download location not seen; Loop-X images are additional real data beyond DeepFluoro Licence checked on 2026-10-10 at https://pmc.ncbi.nlm.nih.gov/articles/PMC10959504/; derived test pairs shareable: unclear; commercial use: unclear. data statement points to the JHU repository (Cloudflare-blocked here, DataCite lists no rights); hip CTs come from NMDID, which has its own data use agreementUse the paired real/simulated images of the same CT and view: pose truth from the simulation geometry
Thorax X-ray and CT interventional 2D/3D datasetInterventional thorax X-ray vs CT, same patientUnverified countLandmarks (archive Landmarks.rar)both imagesdirectCC BY 4.0openwaits for a tool
Its X-ray-to-CT part waits for a 2D/3D pose error; its CT-to-CT part is an example in datasets/examples/PigLungCT, without independent test and development subjects.
Its CT-to-CT part is retained as the PigLungCT example outside the benchmark: all phases belong to one animal. The X-ray-to-CT part waits for a 2D/3D pose metric.
Tomato micro-CT projection datasetCone-beam micro-CT projections of a tomato with geometry metadata (MILabs U-CT)1 object, 720 projections over 360 degreesGeometry metadata fileboth imagesdirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Non-medical soft biological object; corrected projectionsExact by construction after a known rigid move of the reconstruction
TROG 15.01 SPARK (prostate kV intrafraction monitoring)intra-treatment kV and MV projections (2D) to planning CT (3D) of the prostate with implanted gold markers, same patient48 patients x 5 fractions, about 1 million projection imagestumour translation and rotation from KIM marker tracking, plus a ground truth from marker triangulation across projectionsboth imagesdirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
real clinical data; implanted fiducials give an independent truth per projection; very large; marker details not checkeduse triangulated marker positions as truth for markerless 2D/3D (soft tissue or bone) registration of each projection
UltraBones100k (tracked ultrasound and CT of lower limbs)optically tracked 2D B-mode ultrasound frames (2D) to CT bone models of tibia, fibula and foot (3D), cadaver, same specimen14 specimens, about 100,000 tracked framesper-frame probe pose from optical tracking placing each frame in CT space, then intensity-based per-frame refinement; bone-surface labels derived from the CT modelboth imagesdirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
real ex vivo; truth is tracking plus a refinement that itself uses the images, so not fully independent; CT volumes not listed, only STL segmentationsscore registrations against the raw tracked poses (independent) rather than refined ones; add fiducial screws visible in both for a subset
UltraBonesHip (tracked ultrasound and CT of femur and pelvis)optically tracked 2D ultrasound sweeps (2D) to CT bone surfaces of femur and pelvis (3D), cadaver, same specimen5 specimens, 3 anatomies, axial and coronal sweeps of hundreds to 1,500 framesper-frame tracked probe pose expressed in the CT frame (calibration in reconstruct.py); refined poses announcedboth imagesdirectCC BY-NC 4.0non-commercialwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
real ex vivo; segmentation masks and optimised poses 'to be uploaded'; CT volumes not listed, only STLsame as UltraBones100k: keep raw tracking as truth; add a second annotator's bone-surface points on 10 frames per sweep
Utrecht 2D/3D registration gold standard (van de Kraats 2005, spine phantoms)2D X-ray images to 3D CT, MR and 3D rotational X-ray of two spine phantoms2 spine phantoms, several vertebrae, standard start positionsgold standard from the calibrated 3DRX C-arm geometry plus 3D/3D image registrationboth imageshiddenunknownunknownleft out
its images or its truth are not released
original page isi.uu.nl offline; only a 2017 Wayback snapshot; files may be lostask Utrecht (Image Sciences Institute) for the files; otherwise rebuild with a C-arm that records geometry
Utrecht spine 2D-3D gold standard (van de Kraats 2005)Two spine segments: 3D rotational X-ray, CT and MR to 2D X-ray images, with fiducials2 specimens, several 2D views eachFiducial-based gold standard with the standardised 2D-3D evaluation protocol (mTRE, capture range)both imageshiddenunknownunknownleft out
its images or its truth are not released
Former page www.isi.uu.nl/Research/Databases/GS/ returns 404 on 2026-10-10; cited as publicly available in older papersAsk UMC Utrecht for the archive; truth already fiducial-based
UW-ADRC brain dissection photographs with ex vivo MRI (Gazula 2024; BIDS release 2026)Photographs of 4 mm coronal brain slabs to the same specimen's ex vivo FLAIR MRI (photo stack to volume), postmortem28 specimens (24 in the eLife paper), slab stacks at 4, 8 and 12 mm spacingMRI-based SynthSeg labels; manual 9-region tracings on one slice per case by two labellers (eLife 2024)one imagedirectCC BY 4.0 (Zenodo API)openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Whether the ex vivo MRIs are in this record or only in the original eLife release is unverified. Companion: 1,573 slab photos with masks (Zenodo 20647553).Two people click 15 landmarks (ventricle horns, hippocampal head, sulcal fundi) on 10 slab photos and the matching MRI planes.
Veriserum (dual-plane fluoroscopy, knee implant phantoms)Dual-plane X-ray fluoroscopy to 3D implant models (STL and NIfTI), knee implant phantoms (2 femoral, 5 tibial components), same objectabout 110,000 images from 1,600 trials, as dual-plane pairsPose per image from the authors' automatic registration; 200 images with manually registered poses for benchmarking; calibration toolsboth imagesdirectIn Copyright - Non-Commercial Use Permitted (rightsstatements.org InC-NC/1.0)non-commercialwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Implants on phantoms, not bone; most poses are method output; docs at github.com/wjh19990923/VeriserumUse only the 200 manual poses as the test set; a second annotator on 50 of them gives inter-observer pose error
VICTRE in silico breast trial (FDA): phantoms, breastCompress, MC-GPU, TCIA imagesSimulated breast tomosynthesis raw projections of VICTRE digital breast phantoms2 phantoms, 2 projection sets (equal dose; Malmo trial dose profile)Phantoms and acquisition geometry known by construction (geometry file contents unverified)both imagesdirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Synthetic; phantom volumes come from the VICTRE pipeline, not necessarily in this recordRegenerate with VICTRE at recorded geometry plus a known breast deformation for DBT-to-volume truth
Vienna pig head 2D/3D gold standard (Pawiro, Birkfellner 2011)Fresh cadaver pig head with fiducials: CT, MR (T1, T2, PD), CBCT and kV and MV radiographs (Elekta Synergy), same specimen1 specimen, several kV and MV views per 3D modalityFiducial-based gold standard transforms; projection distance error < 2.71 mm, expected TRE < 1.88 mmboth imagesrequestnone stated; 'available from the corresponding author'askneeds a request
the data needs a request or a data use agreement
Only soft-tissue-rich radiotherapy 2D/3D gold standard with MV images; fiducials must be masked Licence checked on 2026-10-10 at https://pmc.ncbi.nlm.nih.gov/articles/PMC3091017/; derived test pairs shareable: unclear; commercial use: unclear. paper: 'The complete data set, which is available from the corresponding author'; ask W. Birkfellner
Wooden block CT and structured light scan (Bossema 2023)FleX-ray micro-CT reconstruction and Artec Spider structured-light surface scan of the same wooden block1 objectNone stated (practice data for a Blender alignment plugin)one imagedirectCC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Non-medical; surface-to-volumeMark 8-10 surface features on both; or apply a known move to the surface scan
X3D (SAX-NeRF sparse-view X-ray benchmark)Simulated cone-beam X-ray projections (TIGRE) of 15 CT volumes (chest, foot, head, jaw, pancreas, objects) with geometry and ground-truth volumes15 scenes, 50 training + test projections each (unverified count)Exact by construction: TIGRE geometry and source CT volumeboth imagesdirectcode MIT; processed data (Google Drive/Baidu) no licence statedaskwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Built for reconstruction; no rigid offset between projections and volume, so a pose task must be made Licence checked on 2026-10-10 at https://github.com/caiyuanhao1998/SAX-NeRF; derived test pairs shareable: unclear; commercial use: unclear. repo LICENSE is MIT, README gives data links without terms; projections are simulated from public CT sets with their own licencesApply known rigid or smooth deformations to the volume and ask for recovery from a few projections
XMAPortal (X-ray Reconstruction of Moving Morphology, public studies)Biplanar high-speed X-ray videos, calibration images and CT scans of animal (and some human) bones for X-ray Reconstruction of Moving MorphologyMany public studies (count unverified), each with trials of synchronised X-ray videosCalibration objects per session; bone poses from marker-based or markerless (scientific rotoscoping) tracking, when uploadedboth imagesdirectnone stated globally; per-study sharing set by the PIaskwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
Legacy portal read-only since June 2026; new portal at xma.xromm.brown.edu; mostly non-human, preclinical/non-medical Licence checked on 2026-10-10 at https://xmaportal.org/webportal/larequest.php?request=documentation; derived test pairs shareable: unclear; commercial use: unclear. portal documentation names no data licence; studies are shared with users or the public by their PI; portal now read-only, moving to xma.xromm.brown.edu; ask each study's PIUse studies with implanted markers: marker-based poses give a 2D/3D bone pose reference for markerless CT-to-video methods
XROMM minipig mastication (XMALab tutorial, XMAPortal)biplanar X-ray video of minipigs chewing and CT of the same animals with tantalum beads in skull and mandible3 pigs, 14 chewing sequences in the study; the public tutorial share holds a subset3D bead positions tracked from both views (about 0.08 mm), used to animate the CT bone modelsboth imagesdirectnone statedaskwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
XMAPortal (xmaportal.org) is now read-only legacy; downloads may need login (unverified). Other public marker-based XROMM studies sit there too. Licence checked on 2026-10-10 at https://xromm.org/projects/pig-feeding/; derived test pairs shareable: unclear; commercial use: unclear. project page describes the study (Metzger, Brainerd) with no data terms; data on XMAPortal, which has no global licence; ask Brown XROMMbeads visible on CT and both X-ray views: bead centroids give the rigid pose per bone per frame; mask the beads when scoring intensity methods
XROMM studies on Zenodo (macaque tongue chewing, opossum mastication, bird shoulder)biplanar X-ray video vs CT bone models with implanted radio-opaque markers, same animalseveral trials per studymarker-based rigid body posesboth imagesdirectCC0 / CC BY 4.0openwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
other deposits: Zenodo 8253287 (Didelphis), 19888511 (ex vivo bird shoulder); related to the XMA Portal studies already listed
xvr-data (DeepFluoro, Femur, Ljubljana repack)CT to X-ray: pelvis (DeepFluoro), femur (Balgrist), cerebral 3D-DSA to 2D-DSA (Ljubljana), same subjectDeepFluoro 6 specimens, Femur 5 specimens / 49 X-rays, Ljubljana 10 patientsOriginal poses and fiducials converted to DICOM/NIfTI with per-image pose files, fiducials and masksboth imagesdirectCard CC BY-NC-ND 4.0; original licences apply (DeepFluoro CC BY-NC 4.0)non-commercialwaits for a tool
2D to 3D: waits for a pose error and a 2D/3D builder
status change for DeepFluoro and Ljubljana 3D-2D-GS-CA: ready-to-use repack in one convention, easiest entry point for 2D/3D

Video · 85

DatasetProblemPairsChecked againstAccessLicenceStatusNotesTo make the truth
ACOUSLIC-AI2D blind-sweep fetal abdominal ultrasound (6 sweeps per case, novice operators, Sierra Leone), frames of the same sweep and between sweeps300 public cases x 6 sweeps (840 frames each)abdomen masks on frames where measurable (optimal/suboptimal), clinician annotatedboth imagesdirectCC BY-NC-SA 4.0non-commercialwaits for a tool
video: waits for point tracking
Masks only on a few frames per sweep; out-of-plane motion dominatesabdomen masks on neighbouring frames as label truth; two people click spine and stomach centres on 20 frame pairs
ANFC_THU nail-fold capillary datasetNail-fold capillaroscopy: 321 images and 219 videos from 68 subjects; video stabilisation, same subject219 videos (raw and stabilised versions)Expert segmentations and keypoints on stills; stabilised videos made by the authors' pipelineone imagerequestAcademic use only, commercial use banned; signed agreement emailed by facultynon-commercialneeds a request
the data needs a request or a data use agreement
Raw vs stabilised video pairs give an algorithm-made reference, not truth.Click capillary apices in first/last frames of 20 videos (two people).
AO-SLO intraframe motion correction data (Azimipour 2018, UC Davis)Adaptive optics SLO frame series of the foveal cone mosaic with fixational eye motion; strip-based registration of raster frames, same eye100 real AO-SLO frames; 200 simulated frames from a synthetic cone mosaic with simulated eye traces (two variants)Simulated set: exact eye-motion traces and motion-free object; real set: noneboth imagesdirectCC BY 4.0 (figshare API)openwaits for a tool
video: waits for point tracking
Also on GitHub rjonnal/intraframe_motion_correction with the simulator. The only public AO-SLO motion set with known motion found; a single real series.Simulator already gives exact truth; for real data, track 20 cones across frames by two people.
Arm reaching ultrasound tissue motion with DUSTrack tracksB-mode ultrasound video of arm tissues during reaching, synchronized with motion capture and EMG36 participantstissue boundary motion from DUSTrack semi-automatic point trackingone imagedirectunverified (figshare)unknownwaits for a tool
video: waits for point tracking
Sci Data 2026. Tracks are algorithm-assisted.two people re-annotate 20 points on 50 frames per participant for 5 participants
BAGLS (Benchmark for Automatic Glottis Segmentation)laryngeal high-speed videoendoscopy, consecutive frames of vocal fold vibration59,250 frames with masks plus videosglottal area masks per frame (frame-wise)both imagesdirectCC BY-NC-SA 4.0non-commercialwaits for a tool
video: waits for point tracking
Glottis opening is a topology change, not a smooth warp; vocal fold edges are the trackable structuretwo people click vocal process and anterior commissure on 50 frame pairs
BEDLUS (Boston Emergency Department Lung Ultrasound)2D lung ultrasound videos, 113 patients, frames of the same video1,419 videos15,755 B-line annotations on 10,371 frames (artefact lines, not anatomy)both imagesdirectcustom terms specific to this dataset; restricted accessaskwaits for a tool
video: waits for point tracking
Fails as anatomy truth: B-lines are reverberation artefacts that do not move with tissue; pleura and ribs are the only trackable structures Licence checked on 2026-10-10 at https://doi.org/10.7910/DVN/GLCZRB; derived test pairs shareable: unclear; commercial use: unclear. Harvard Dataverse (DataCite rights: restrictedAccess, 'Custom terms specific to this dataset'); README: download 'after requesting access'; terms text not readable without the guestbooktwo people mark rib shadows and pleural line points on 20 frame pairs
Breast lesion ultrasound video dataset (CVA-Net, MICCAI 2022)2D breast ultrasound sweep videos through a lesion, frames of the same video188 videos, 25,272 frames (28-413 per video)lesion bounding box on every frame by two pathologists (per paper)both imagesdirectnon-commercial research or education use only (repo text)non-commercialwaits for a tool
video: waits for point tracking
Sweep through the lesion, so frames differ out of plane; boxes are weak truth; Baidu or Google Drive downloadbox centres as weak correspondences on frames a few apart; two people click 5 points on 20 nearby frame pairs
Brno video-ophthalmoscope retinal sequencesLow-cost video-ophthalmoscope fundus sequences, frame to frame, same eye23 sequences, 5,840 framesRegistration error measured by authors (method of truth unverified)hiddennot released (no data statement seen)unknownleft out
its images or its truth are not released
Ask Radim Kolar (Brno).Not released: ask the authors; if obtained, two people click 5 vessel bifurcations on 20 frame pairs per sequence.
Bronchoscopy visual odometry dataset (Deng et al., IROS 2023)Bronchoscopy video in an airway phantom (and ex vivo human lungs in the paper) with reference camera trajectoriesSeveral phantom sequences at difficulty levels (count unverified)Reference trajectories per frame (source of poses unverified, likely EM or robot tracking)both imagesdirectrepo GPL-3.0; data (Google Drive) no licence statedaskwaits for a tool
video: waits for point tracking
Only phantom data public; no 3D airway model shipped (per BREATH-VL), so not 2D/3D until a phantom CT is added Licence checked on 2026-10-10 at https://github.com/SIRGLab/bronchoscopy-dataset-and-benchmark; derived test pairs shareable: unclear; commercial use: unclear. README gives only a Google Drive link; the GPL-3.0 LICENSE reads as a code licenceScan the phantom (CT) and register its model to the trajectory frame with 6+ fiducial points; then frames become 2D/3D pairs
C3VD-DEFCOL (deformable colonoscopy, C3VDv3)Rendered colonoscopy clips of C3VD/C3VDv2 colon meshes under controlled peristalsis and centreline motion, with sim-to-real texture translationunverified count of deformable clipsPer-frame depth, normals, flow, camera poses and time-stamped deformed meshes, exact by constructionboth imagesdirectCC BY 4.0openwaits for a tool
video: waits for point tracking
Synthetic deformation of real phantom geometry; RGB is generated by a video model, so appearance is not physical Licence checked on 2026-10-10 at https://github.com/DurrLab/C3VDv3; derived test pairs shareable: yes; commercial use: yes. README 'License: This work is licensed under CC BY 4.0'; sim2real subfolder has its own licence
C3VDRegColonoscopy depth point clouds vs CT-mesh raycasts at matched poses (derived from C3VD)10,015 partial-to-partial pairs (2,088 test)Camera poses from C3VD 2D-3D registrationboth imagesdirectCC BY-NC-SA 4.0 (C3VD-Raycasting-10k, UCL RDR)non-commercialwaits for a tool
video: waits for point tracking
status change for C3VD: new registration benchmark; point clouds, not images Licence checked on 2026-10-10 at https://doi.org/10.5522/04/30640043; derived test pairs shareable: yes; commercial use: no. paper: full dataset on UCL RDR, and 'Redistribution of derived point clouds should remain consistent with the upstream C3VD terms'; share-alike
CADICA (invasive coronary angiography videos)Coronary angiography videos (Siemens Artis Zee, 10 fps), Malaga; frames of the same run and several views per patient668 videos from 42 patients (382 selected), 6,126 keyframesLesion boxes with stenosis grade on keyframes, by cardiologistsone imagedirectCC BY 4.0openwaits for a tool
video: waits for point tracking
Full runs include pre-contrast frames: mask/live pairs for DSA-style motion correction.Pick pre-contrast and contrast frames at the same cardiac phase; two people click 15 background landmarks (rib edges, catheter tip, diaphragm points) on 30 pairs
Capillary Dataset (nail-fold capillaroscopy, diabetes)Nail-fold capillaroscopy images and videos at 390x; frame-to-frame stabilisation and mosaicking, same subject202 subjects, 6,695 images plus a video folderMorphology bounding boxes and classes on stills; no motion truthone imagedirectApache-2.0 (HF card)openwaits for a tool
video: waits for point tracking
Hand micro-shake makes video stabilisation a real task (projection registration papers use private videos).Simulate known shifts and small warps on stable frames; or click 10 capillary apex points across 2 frames of 20 videos.
CardiacUDA (cardiacUDC)2D echo video, four views (only A4C on Kaggle per HF mirror), two Chinese hospitals, frames of the same video992 videos in paper; 364 A4C videos released (293 with masks); val/test videos masked on every framechamber masks (LV, RV, LA, RA) by 5-6 physicians; every frame in val/test, 5 frames per training videoboth imagesaccountCC BY-NC 4.0non-commercialneeds an account
the download needs an account
Per-frame four-chamber masks give dense label truth across a cycle; Kaggle: kaggle.com/datasets/xiaoweixumedicalai/cardiacudc-dataset Licence checked on 2026-10-10 at https://www.kaggle.com/datasets/xiaoweixumedicalai/cardiacudc-dataset; derived test pairs shareable: yes; commercial use: no. authors' Kaggle upload (Kaggle API licenseName); GitHub repo has no licence; Kaggle download needs an accounttwo people click valve hinge points and apex on ED/ES of 20 test videos
CARDINAL hypertensive echocardiography (Ling 2023)2D apical 2CH and 4CH echo of 239 hypertensive patientsSequences from 239 patientsnnU-Net pseudo-labels validated by an expert cardiologistone imagerequestprivate datasetaskneeds a request
the data needs a request or a data use agreement
Access route not found (unverified); FIMH 2023 paper Licence checked on 2026-10-10 at https://arxiv.org/abs/2305.01997; derived test pairs shareable: no; commercial use: no. FIMH 2023 paper: 'We introduce a new private dataset, named CARDINAL' (240 patients, Hospices Civils de Lyon); ask CREATISSimulate from it with a known motion (as TAS-1K does), or experts track 10 myocardial points per sequence
CathAction (endovascular X-ray fluoroscopy videos)X-ray fluoroscopy videos of catheterisation (subjects unverified); frame-to-frame motion and device trackingabout 500,000 annotated frames; 25,000 catheter and guidewire masksCatheter/guidewire masks and action labels, by the authors; no anatomical correspondencesboth imagesrequestCC BY-NC-SA (download via form)non-commercialneeds a request
the data needs a request or a data use agreement
Device masks move with the device, not the anatomy: weak as registration truth.Click 10 vessel or bone landmarks on frame pairs 1 s apart in 20 clips (two people) for respiratory/cardiac motion TRE.
Cholec8080 laparoscopic cholecystectomy videos with phase and tool labelsFrame pairs from 80 videosNone for motion; phase and tool presence labelsone imagedirectCC-BY-NC-SA 4.0non-commercialwaits for a tool
video: waits for point tracking
Fails the truth criterion; truth added only by Endo-TAPC5 labellingTwo people label 20 tissue points on first and last frames of 50 clips, STIR-style
CLiMB 2026 (EndoVis colonoscopy localisation and mapping)real colonoscopy video (curated EndoMapper clips), camera trajectory and map from monocular videoshort test clips cut from about 10-12 source sequences; training data unannotatedCOLMAP pseudo ground-truth poses, visually checked, scaled to an assumed 50 mm colon diameter; kept privaterequestEndoMapper terms: research use only, no redistribution, statement of intended useaskneeds a request
the data needs a request or a data use agreement
SfM pseudo-GT is circular for SfM-based methods; only locally rigid clips kept.a colon phantom with known geometry (C3VD style) for metric truth; on real clips two people mark 10 matched mucosal features across frames
CLUST (Challenge on Liver Ultrasound Tracking, MICCAI 2014/2015)2D+t and 3D+t liver ultrasound, free breathing, same subject over time (track vessel landmarks frame to frame)86 sequences (64 2D+t, 22 3D+t) from 5 scanners; 1-5 landmarks per sequence, annotated on about 10-13% of framesvessel landmark positions, mean of 3 expert observers; training annotations given, test annotations held by organisers (first frame only given)both imagesrequestchallenge rules: data used solely for participating in the CLUST challengeaskneeds a request
the data needs a request or a data use agreement
Site says the challenge is moving institute and cannot process access or evaluation requests now. A 9-sequence volunteer subset without labels is a direct download (8.6 GB): data.vision.ee.ethz.ch/cvl/datasets_extra/usliverseq.zip Licence checked on 2026-10-10 at https://caim.research.it.uu.se/clust/data.html; derived test pairs shareable: no; commercial use: no. login issued on request from an institutional email; rules limit use to the challenge, so benchmark use needs the organisers' permission
Colon10Kundistorted colonoscopy frames from 20 subsequences, frames of the same place10,126 frames; query frames with labelled positive intervalsmanual same-place labels (frame intervals), no point correspondences; camera intrinsics givenone imagedirectcopyright UNC Medical School; no licenceaskwaits for a tool
video: waits for point tracking
Low resolution (270x216); place labels give pairs to register but no geometric truth Licence checked on 2026-10-10 at https://endoscopography.web.unc.edu/place-recognition-in-colonoscopy/; derived test pairs shareable: unclear; commercial use: unclear. page: 'Data copyright belongs to UNC Medical School'; no reuse termsCOLMAP on each subsequence for relative poses; two people click 8 fold or vessel points on 30 positive pairs
Common Carotid Artery Ultrasound Images (Momot, Mendeley)2D longitudinal carotid B-mode time series, 11 subjects, frames of the same series1,100 frames (100 per subject)artery masks by a technician, verified by an expertboth imagesdirectCC BY 4.0 (Mendeley default, unverified)openwaits for a tool
video: waits for point tracking
Consecutive frames with masks: weak label truth for wall motiontwo people track 3 wall points (near wall, far wall, plaque) over the series, as Zahnd 2013 did
CoronaryDominanceMulti-view coronary angiography videos of 1,574 patients (dominance labels)1,574 patients, many videos each (160,320 extracted frames per XA-170K)Coronary dominance class per patient; no spatial marksone imagedirectCC0 1.0openwaits for a tool
video: waits for point tracking
Largest open angiography video source; frames of the same run give real cardiac/breathing motion.Frame pairs at end-diastole of consecutive beats; bifurcations clicked by two people on 30 runs
D4D (Dresden 4D reconstruction of non-rigid abdominal surgical scenes)da Vinci Xi stereo endoscope video of porcine cadaver abdomen with structured-light (Zivid) 3D geometry at clip start and end, optically tracked6 porcine sessions; whole-deformation, incremental and moved-camera clipsStructured-light point clouds and curated camera poses registered by optical tracking plus manual ICP-style alignmentboth imagesdirectCC BY 4.0 (paper); dataset licence on OPARA unverifiedopenwaits for a tool
video: waits for point tracking
3D side is structured light, not CT; video-to-surface registration under deformationadd tissue tattoos/beads in future captures; meanwhile surface distance to the end scan as the check
DL_Track_US example and test data2D ultrasound images and videos of lower-limb muscles (gastrocnemius, vastus lateralis), frames of the same videoexample videos plus test images (Zenodo 7866598 older copy, 1.1 GB)manual fascicle length, pennation angle and thickness on test frames (per paper); three manual observers on one video in UltraTimTrack paperboth imagesdirectCC BY 4.0 (Zenodo copy)openwaits for a tool
video: waits for point tracking
Measurements, not point correspondences; OSF content not opened (unverified)fascicle end points tracked by two people on 50 frames give point truth
DRENDS (depth in robotic endoscopy with dynamic scenes)stereo robotic laparoscopy of a phantom and ex vivo porcine tissue under manipulation and non-rigid deformationsequences over several organs and three illuminations (count unverified)ground-truth point cloud per frame and calibration; no correspondencesone imagedirectCC BY 4.0openwaits for a tool
video: waits for point tracking
Sci Data 2026. Depth truth only; frame-to-frame motion must be derived.pins or ink fiducials on the tissue (as in SuPer) for tracks; scene flow from consecutive point clouds checked by a person
EchoNet-LVH2D echo video, parasternal long axis, adults, frames of the same videoabout 12,000 videos (paper says 23,212 annotated videos)expert keypoints for IVS, LVID and LVPW measurements on ED and ES frames (4 points per frame line)both imagesaccountStanford University Research Use Agreement (terms text not visible)non-commercialneeds an account
the download needs an account
Keypoints on two frames per video: usable as sparse ED-ES landmark truth; count differs between AIMI page and paper
EchoNet-Pediatric2D echo video (apical 4-chamber and parasternal short axis per the paper; page does not list views), children 0-18 y, frames of the same video7,643 videos; ED-ES frame pairsexpert LV tracings and measurements on ED and ES frames (page says 'measurements, tracings, calculations')both imagesaccountStanford University Research Use Agreement (terms text not visible on page); commercial licence for a feenon-commercialneeds an account
the download needs an account
Label truth only (LV masks at two frames), no point correspondences; views per paper, unverified on page; download via RedivisLV contour points resampled by arc length give weak correspondences; for points, two experts click 6-10 myocardial landmarks (annulus, apex) on ED and ES of 30 test videos
EchoTracker in-house myocardial trajectory sets (NTNU/St. Olavs)2D echo videos with myocardial point trajectories over a cardiac cycle, LV and RVTraining 643 patients / 1,922 videos; test 371 / 1,112; RV out-of-distribution 701-1,254 patientsTrajectories from a traditional tracker refined by clinical expertsboth imageshiddenNot statedunknownleft out
its images or its truth are not released
Private (GE Vivid E95/E9)
EchoXFlow (beamspace echocardiography, Akershus)Clinical echo in native beamspace: 2D/3D B-mode, tissue Doppler, colour and spectral Doppler with ECG; same patient over time37,125 recordings from 666 exams (622 patients)Dense speckle-tracking myocardial contours (36,132 frames) and 3D LV meshes (10,525 volumes) from routine reporting; tissue Doppler velocitiesboth imagesdirectCC BY-NC-SA 4.0non-commercialwaits for a tool
video: waits for point tracking
Tissue Doppler gives an independent velocity check for motion estimatesTissue Doppler velocity at sample points as an independent motion truth; experts check speckle-tracking contours
Endo-Mix6 (EndoMatcher matching pairs)endoscopic frame pairs across colonoscopy, laparoscopy, synthetic phantoms (C3VD, EndoSLAM, SCARED, EndoMapper, Colonoscopic); bronchoscopy domain private~1.2M pairs across six domains (five public)correspondence labels from SfM and simulated transformations (pseudo truth)both imagesdirectnone stated; derived from source datasets with their own licencesaskwaits for a tool
video: waits for point tracking
Derived pairs from listed sources; Baidu Netdisk download; SfM labels are not independent truth Licence checked on 2026-10-10 at https://github.com/Beryl2000/EndoMatcher; derived test pairs shareable: unclear; commercial use: unclear. repo has no licence; README points to the source datasets and a Baidu downloadverify a subset by clicking 10 points on 50 pairs per domain
Endo-TAPC5 endoscopic point-tracking benchmarkEndoscopic clips from Hamlyn, Cholec80 and 40 private hospital videos; tissue point trackingNot statedPoints labelled to the STIR standard by the authorsboth imageshiddenNot statedunknownleft out
its images or its truth are not released
Release promised, not yet out; 40 videos private
EndoAbS (endoscopic abdominal stereo phantom)stereo endoscope images of phantom abdominal organs under light, smoke and distance changes, left-right and across conditions120 stereo pairslaser-scanner surface point cloud in the left camera frame (0.43 mm camera-laser calibration error), stereo calibrationboth imagesdirectnone stated; Zenodo files restricted (access on request)askwaits for a tool
video: waits for point tracking
Static phantom: the scan plus calibration gives exact correspondences between left, right and repeated views Licence checked on 2026-10-10 at https://zenodo.org/records/640084; derived test pairs shareable: unclear; commercial use: unclear. Zenodo record public, files restricted, no licence or access conditions shown; request access from the authors (Penza, De Momi)
EndoMapper (calibrated complete endoscopy procedures)full calibrated colonoscopy and gastroscopy videos (HD), frames of the same procedure; same-patient repeat procedures; simulated deforming colon sequences96 real sequences (~24 h); COLMAP clusters with covisible frame lists; a few simulated sequencesreal: COLMAP up-to-scale poses and sparse 3D points (pseudo truth, organiser-checked for SimCol3D task 3); simulated: exact depth, camera trajectory and deformationboth imagesrequestresearch use only, no redistribution, statement of intended use required (Synapse conditions)non-commercialneeds a request
the data needs a request or a data use agreement
Request approvals reported slow on the Synapse forum; COLMAP covisibility gives sparse correspondences between framesCOLMAP points reprojected give sparse correspondences; two people click 10 vessel crossings on 20 frame pairs to check them
EndoNeRF da Vinci prostatectomy clips (Wang 2022)Stereo endoscopic clips of deforming tissue (pulling, cutting) for deformable reconstructionSample clips (count not stated on the page)None: stereo depth (STTR-light) and manual tool masksone imagedirectnone statedaskwaits for a tool
video: waits for point tracking
Fails the truth criterion; Google Drive link Licence checked on 2026-10-10 at https://github.com/med-air/EndoNeRF; derived test pairs shareable: unclear; commercial use: unclear. README links a 'Sample Dataset' on Google Drive (earlier via a Google Form); repo has no licenceTwo people track 20 tissue points across each clip; or tattoo-style IR markers as in STIR for new recordings
EndoSLAMendoscope and capsule video in ex vivo porcine GI organs and a silicone colon phantom with CT35 ex vivo sub-datasets with 6D poses, phantom colonoscopy, synthetic setsrobot-arm 6D poses, CT of the phantom, 3D scansboth imagesdirectCC BY 4.0openwaits for a tool
video: waits for point tracking
Video-to-CT registration possible on the phantom part.
EVEREST PCV ICGA and FA sequences (Tsai 2009/2016)Heidelberg confocal SLO ICGA dynamic sequences with IR and FA frames, polypoidal choroidal vasculopathy, same eye60 sequences of about 20 frames (EVEREST); 62 ICG sequences (Taipei)reading-centre polyp ground truth; authors report 1.58 px mean alignment error from their own checkone imagehiddentrial data, not releasedaskleft out
its images or its truth are not released
dye filling changes appearance across frames: a hard intra-sequence multimodal casechoroidal and retinal vessel bifurcations on early and late frames by a program, verified by a person
Fetoscopy placenta vessel dataset (Bano 2020, UCL)In vivo TTTS fetoscopic videos for vessel segmentation and mosaicking; consecutive frames6 in vivo clips (up to 400 frames); 483 annotated framesVessel masks verified by a fetal medicine specialist; no registration ground truthone imagedirectCC BY-NC-SA 4.0non-commercialwaits for a tool
video: waits for point tracking
Predecessor of FetReg 2021; fails the truth criterion Licence checked on 2026-10-10 at https://doi.org/10.5522/04/30536171; derived test pairs shareable: yes; commercial use: no. UCL RDR record 'Fetoscopy Placenta Dataset' (DataCite rights); UCL web page returned 403Vessel junctions found by a program then verified by a person on frame pairs 5-10 apart
FetReg 2021 (fetoscopy segmentation and registration)Fetoscopic video frames, mosaicking (frame-to-frame homography)24 surgeries: 2,718 annotated frames + 24 clips (9,616 frames)Vessel/fetus/tool masks; no registration ground truth (mosaic consistency metric)one imageaccountCC BY-NC-SA 4.0non-commercialneeds an account
the download needs an account
Fails truth criterion for registration Licence checked on 2026-10-10 at https://doi.org/10.5522/04/30417166; derived test pairs shareable: yes; commercial use: no. UCL RDR record 'FetReg: Largescale Multi-centre Fetoscopy Placenta Dataset' (DataCite rights); no Synapse account needed any moreTwo people click 6 vessel points on 20 frame pairs per clip to score homographies; vessel masks on both frames as a weaker check.
FI3S From Images to 3D Shapes (Kaggle)synthetic RGB-D recordings of deforming objects, frame to framemany sequences (count unverified)synthetic registration ground truth with depthboth imagesaccountCC BY-SA 4.0openneeds an account
the download needs an account
synthetic
FIDO (MICCAI 2026)Listed with keywords calibration, localization, registration, trackingUnverifiedUnverifiedhiddenUnverifiedunknownleft out
its images or its truth are not released
Only the MICCAI registered-challenge entry seen; data unverifiedData unknown: decide after release.
Gastrocnemius fascicle tracking data (Drazan 2019)2D B-mode ultrasound of medial gastrocnemius during maximal contractions, frames of the same video5 subjects of trials plus a validation setmanually tracked fascicles in validation data (fascicle end points per frame, per paper)both imagesdirectCC BY 4.0openwaits for a tool
video: waits for point tracking
Truth is fascicle line end points, sparse; check validation_data.zip content (unverified)two people track 3 fascicle end points and aponeurosis points on 50 frames
Hamlyn Centre laparoscopic/endoscopic video datasetsin vivo stereo laparoscopy, beating-heart phantom stereo video with CT, GI endoscopy region trackingseveral sequences (2 phantom heart sequences with CT; 10 in vivo GI tracking sequences)CT surface for the phantom heart; region tracking labels for GI sequencesdirectpermission statement: 'Permission is given to use and publish all data on this website', cite the papersopenwaits for a tool
video: waits for point tracking
Site refused connection on 2026-10-10; details from search snippets. Used in 1 of the 2D and microscopy registration papers read (Endo-TTAP 2025 (via Endo-TAPC5)). Licence checked on 2026-10-10 at https://web.archive.org/web/20251224171356/https://hamlyn.doc.ic.ac.uk//vision/; derived test pairs shareable: yes; commercial use: yes. a blanket permission, not a named licence, and silent on commercial use (read via Wayback, live site timed out); HF mirrors relabel parts as GPL-3.0Phantom heart sequences have a CT surface: two people click 10 surface points in stereo frames and on the CT; in vivo, two people track 10 points over 20 frames.
HMC-QU2D echo A4C and A2C video, myocardial infarction patients, frames of one cardiac cycle162 A4C and 160 A2C recordings; 109 A4C with masks on every frame of one cycleLV wall masks on every frame (109 A4C videos), Qatar University / Hamad Medicalboth imagesaccountCC BY-NC-SA 3.0 IGOnon-commercialneeds an account
the download needs an account
Per-frame wall masks over a full cycle: label truth for frame-to-frame registration; low image quality in part Licence checked on 2026-10-10 at https://www.kaggle.com/datasets/aysendegerli/hmcqu-dataset; derived test pairs shareable: yes; commercial use: no. Kaggle dataset licence (Kaggle API licenseName); share-alike; Kaggle download needs an accounttwo people click annulus and apex points on ED and ES of 20 videos
HUNT and SSHF test-retest echo strain sets (NTNU)Repeated apical 2CH/3CH/4CH acquisitions of the same patients; strain reproducibilityHUNT 30 patients and SSHF 40 patients, 2 acquisitions per view; HUNT-A 88 patients, 2 expertsExpert speckle-tracking reference points and GLS (EchoPAC)both imageshiddenNot statedunknownleft out
its images or its truth are not released
Private Norwegian cohorts
HyperKvasir (GI endoscopy images and videos)gastroscopy and colonoscopy videos and still images, consecutive framesvideos and 110k images (counts per OSF, unverified)image class labels and polyp masks on a subset; none spatial across framesone imagedirectCC BY 4.0 (per data paper, unverified)openwaits for a tool
video: waits for point tracking
Fails truth; large open pool of GI video for building frame-pair benchmarksCOLMAP on short clips for sparse correspondences; two people click 8 points on 30 frame pairs
iMED 2026 (EndoVis multi-endoscope pose and deformable view synthesis)two synchronized endoscopes viewing ex vivo organs, human cadavers and one live pig, with tool-tissue interaction and breathing340 sequences (105 for pose estimation, 113 for deformable view synthesis)relative pose between the two endoscopes from ArUco markers fixed to the anatomy, provided for all frames (markers then inpainted)both imagesdirectCC BY (challenge design document)openwaits for a tool
video: waits for point tracking
Design document says the training set is public; download page unverified. Test ground truth hidden.
Kvasir-Capsulevideo capsule endoscopy (Olympus), consecutive frames of the small bowel at 2 fps117 videos (~4.7M frames); 47,238 labelled framesfinding labels and bounding boxes onlyone imagedirectCC BY 4.0 (OSF/Sci Data); simula page asks permission for competitions/commercial useopenwaits for a tool
video: waits for point tracking
Fails truth; low frame rate and large inter-frame motion; licence statements differ between pagestwo people click 6-8 mucosal landmarks on 50 consecutive pairs; boxes of the same lesion across frames as weak check
Marmoset reaching, XROMM vs DeepLabCut (Dryad)biplanar X-ray video and two RGB cameras of marmosets reaching, in one calibrated frame; subcutaneous tantalum beads2 animals, 22 reaches11 bead tracks from XMALab (about 0.06 mm), DLC/Anipose tracks, camera calibrationsone imagedirectCC0openwaits for a tool
video: waits for point tracking
Fails image-to-image registration: it is keypoint tracking; no CT confirmed.calibration gives exact X-ray-to-RGB camera geometry; projecting bead 3D positions into RGB gives point truth for cross-modal video registration (beads not visible in RGB)
MIMIC-IV-ECHOechocardiography DICOM loops (TTE, stress, TEE) with structured measurements, same patient across studies and frames within loopsstructured measurements from 206,488 studies of 91,372 patients; DICOM subsetcardiologist measurements (chamber sizes, function) only, no tracings in imagesone imagecontrolledPhysioNet Credentialed Health Data License 1.5.0non-commercialneeds a request
the data needs a request or a data use agreement
Fails spatial truth; a large pool to build an echo benchmark from, with repeat studies of the same patienttwo experts click annulus and apex on ED/ES of 50 loops; public LV segmenter (EchoNet) for label checks
MOSXAV (X-ray angiography video segmentation)Coronary X-ray angiography cine runs (512x512, 33-70 frames); frames of the same run (cardiac and breathing motion)42 sequences: 30 train/val (2,335 frames, masks every 5 frames), 12 test (488 frames, every frame)Radiologist instance masks per vessel branch, catheters and balloons, tracked across framesboth imagesdirectCC BY-NC-SA 4.0non-commercialwaits for a tool
video: waits for point tracking
Per-frame masks of the same catheter and vessel branches give frame-to-frame correspondence (mask overlap after warping); opacification changes over time. Licence checked on 2026-10-10 at https://github.com/xilin-x/MOSXAV; derived test pairs shareable: yes; commercial use: no. README section 5 and LICENSE file; share-alike, non-commercialUse catheter masks (no contrast change) and branch bifurcations clicked by two people on 20 frame pairs per run
MPI Sintel, KITTI flow, Springoptical flow on natural and rendered videothousandsdense flow, hidden for the test setsboth imageshiddenvariousunknownleft out
its images or its truth are not released
Outside medicine: uploads of fields, capped (Spring: 3 per 30 days).
MTJtrack (gastrocnemius muscle-tendon junction, DeepLabCut)2D ultrasound cine of the medial gastrocnemius muscle-tendon junction during walking and contractions, 15 subjects3,482 manually tracked frames across subjectsmanual MTJ point position per frame (single point)both imagesdirectnone statedaskwaits for a tool
video: waits for point tracking
One landmark per frame: thin truth; images ship inside DeepLabCut projects (check) Licence checked on 2026-10-10 at https://github.com/rlkrup/MTJtrack; derived test pairs shareable: unclear; commercial use: unclear. repo holds DeepLabCut projects (labelled frames and trained networks) with no LICENSEadd 3-5 more points (aponeurosis ends, fascicle insertions) by two people on 50 frames
MV-dVRK (multi-viewpoint ex vivo surgical benchmark)several exposure-synchronized da Vinci stereo endoscopes viewing ex vivo tissue; static and dynamic scenesstatic scenes plus 10 dynamic sequences with deformationSfM reference geometry validated against an industrial 3D scanner, ground-truth camera posesboth imagesdirectunverifiedunknownwaits for a tool
video: waits for point tracking
Project page linked from arXiv, not opened.for the dynamic sequences add tissue fiducials to get tracks
PACMAN-AMI NIRS-IVUS and OCT pullbacks (IVUS-OCT co-registration)Coronary intravascular pullbacks: NIRS-IVUS and OCT of the same vessel; longitudinal and rotational co-registration of frame sequences, same patient230 patients, 714 vessels (paired pullbacks)Expert lumen borders (about 62k frames per modality), side branches and calcium on 10k frames each; method tested against analysts on 22 patients (77 vessels), details unverifiedboth imageshiddennot public (licensing restrictions stated in IntraCross repo)unknownleft out
its images or its truth are not released
No public intravascular pullback pair set found; baseline/follow-up IVUS registration code exists (dmolony3/IVUS) but data are private.Side-branch ostia matched by two analysts across modalities or time points give longitudinal and angular landmarks.
pCLE laryngeal frame-pair landmarks (Wuerzburg/Erlangen 2026)Probe-based confocal laser endomicroscopy video frames of the larynx; frame-to-frame non-rigid mosaicking, same patient132 frame pairs from 14 sequences, 4 patientsAbout 10 manual cell-corner landmark pairs per frame pair, BigWarp, single non-physician annotatorboth imageshiddennot stated; paper promises code, not dataunknownleft out
its images or its truth are not released
Only clinical pCLE set with landmark TRE; ask the authors (Breininger lab) for the data.Second annotator on the same pairs to measure inter-rater error.
Retinal microsurgery in-vivo tool tracking (Sznitman 2012)In-vivo retinal surgery microscope video, frame to frame18 sequences x 200 frames (1920x1080) (per third-party listing)Instrument joint annotations onlyone imagerequestnone stated; 'publicly available online via the corresponding authors website'askneeds a request
the data needs a request or a data use agreement
Retinal background could serve frame-to-frame mosaicking; no anatomical truth. Fails truth. Licence checked on 2026-10-10 at https://infoscience.epfl.ch/record/180491/files/paper577.pdf?version=1; derived test pairs shareable: unclear; commercial use: unclear. paper points to sites.google.com/site/sznitr/, which now redirects to a Google sign-in; ask R. SznitmanVessel bifurcations tracked by a program and verified by a person on 10 frames per sequence.
Robot-tracked stereo phantom and chicken tissue recordings (TU Hamburg 2025)Stereo endoscope video of a 3D-printed phantom and chicken breast moved by a robot arm; 3D tissue tracking30 s trajectories, 3 per speed (10-80 mm/s), 1080p 30 fpsRobot end-effector path (ABB IRB-120) after camera-robot calibration with 125 checkerboard posesboth imageshiddenNot statedunknownleft out
its images or its truth are not released
Private; rigid motion only, so a limited check
RVD (smartphone fundus video dataset)Handheld smartphone fundus videos (25 fps, 2-30 s), frame to frame, same eye635 videos, >130,000 framesVessel and A/V masks on selected frames; pulse-width frames at the discone imagedirectCC BY-NC-ND 4.0non-commercialwaits for a tool
video: waits for point tracking
Frame registration/mosaicking with motion blur; truth is per-frame masks, not correspondences. ND licence limits derived sets.Bifurcations tracked by a program across frames and verified by a person on 5 frames per video for 30 videos.
SCARED (EndoVis 2019)stereo endoscopic video of porcine cadaver abdomen with structured-light depth keyframes9 datasets (7 train, 2 test); 35 reliable keyframesstructured-light depth at keyframes; robot-kinematics poses (inaccurate) for other framesboth imagesrequestnon-commercial academic research (per SCARED-C card)non-commercialneeds a request
the data needs a request or a data use agreement
Downloads page is restricted to challenge members.
SCARED-CSCARED frames with re-estimated camera poses and reprojected depth17,135 RGB-D frames in 12 keyframe sequencesCOLMAP poses scaled to structured-light depth (authors)both imagesdirectscared-c-research-only (inherits SCARED, non-commercial)non-commercialwaits for a tool
video: waits for point tracking
Card asks users to verify access rights to original SCARED; truth partly algorithmic.
SEE-AI Project capsule datasetsmall-bowel capsule endoscopy stills (PillCam SB3) from 523 videos18,481 images; frame adjacency not givenlesion boxes (YOLO)one imageaccountCC BY 4.0openneeds an account
the download needs an account
Fails pairing: stills without sequence orderonly by construction (known warp)
Semantic-SuPer (dVRK ex vivo tissue tracking with pins)stereo dVRK video of chicken over beef being pushed and grasped (ex vivo), frames of the same trial4 trials x 150 framestrajectories of 30-60 green pins on the tissue (colour-detected physical markers), robot kinematics, semantic masksboth imagesdirectnone statedaskwaits for a tool
video: waits for point tracking
Distinct from SuPer (listed); pins are visible markers, so appearance is not marker-free Licence checked on 2026-10-10 at https://github.com/ucsdarclab/Python-SuPer; derived test pairs shareable: unclear; commercial use: unclear. README gives a Google Drive data link; repo has no LICENSE
SERV-CTstereo endoscope images to cone-beam CT of ex vivo porcine abdomen; endoscope pose in CT16 stereo pairs, 2 specimensCT-derived depth and disparity after constrained manual alignment of endoscope and anatomy in CTboth imagesdirectCC BY 4.0openwaits for a tool
video: waits for point tracking
Depth benchmark; CT-to-video pose is recoverable.
SLIM slit-lamp retinal mosaicing sequences (Quantel/Clermont)Slit-lamp retinal video sequences through a contact lens, long sequential mosaics, plus SLIM-to-FA registrationsequences from 11 patientsNone releasedhiddennot releasedunknownleft out
its images or its truth are not released
Industrial data (Quantel Medical).Not released (industrial data): no truth possible.
Soleus fascicle ultrasound for real-time tracking (Zenodo 4895233)2D B-mode soleus ultrasound at 60 fps, 6 subjects x 5 ankle tasks incl. walking, frames of the same video30 videos of ~24 s (~1,440 frames each)none in the record (UltraTrack .mat format for semi-automatic tracking)one imagedirectCC0 1.0openwaits for a tool
video: waits for point tracking
Fails truth; CC0 makes it easy to redistributetwo people track 3 fascicles every 10th frame on 5 videos; or known warp of real frames
Speech real-time MRI, 75 speakers (USC)2D real-time MRI video and 3D volumes of the vocal tract, same speaker75 speakersnoneone imagedirectCC0 (unverified)openwaits for a tool
video: waits for point tracking
two people trace tongue and palate on 50 frames per speaker
SpineAlign (in vivo RGB-D point clouds of open spine surgery with CT/MR vertebra models)Intraoperative colour point cloud sequences (RGB-D) of lumbar open spine surgery to preoperative CT or MR vertebra meshes, same patient (Humanitas)24 surgeries (20 train, 4 test sequences)Coarse pre-to-intraoperative alignment from surgeon-identified anatomical landmarks; proxy labels from CTone imagedirectCC BY 4.0openwaits for a tool
video: waits for point tracking
Authors of End2Reg state no reliable ground truth exists; in vivo complement to SpineDepthTwo surgeons place 4 landmarks per vertebra (spinous, laminae, transverse processes) on point cloud and mesh to give TRE
SpineDepthRGB-D video of pedicle screw placement to CT-derived vertebra models, 10 cadaveric lumbar spines10 specimens, about 318,000 frames from 2 viewpointsoptically tracked pose of each vertebra L1-L5 per frame, CT vertebra meshesboth imagesdirectnone stated; 'Open Access' download pageaskwaits for a tool
video: waits for point tracking
Surface/point-cloud to CT registration, not image-to-image. Licence checked on 2026-10-10 at https://rocs.balgrist.ch/en/open-access/spinedepth/; derived test pairs shareable: unclear; commercial use: unclear. Balgrist ROCS page and paper data statement (J Imaging 2021) give free downloads with no licence; HF mirror CC BY-NC 4.0 is a third party's choice
StereoMISin vivo porcine stereo laparoscopy (da Vinci Xi) with breathing, tool motion and tissue deformation, frames of the same sequence11 sequences from 3 pigscamera forward kinematics synchronised with video (camera pose), automatic tool masks; no tissue correspondencesone imagedirectCC BY-NC-SA 4.0non-commercialwaits for a tool
video: waits for point tracking
Kinematics give camera motion only; tissue moves with breathing, so pose does not give pixel truthstereo depth plus kinematic pose gives correspondences in static stretches; two people click 15 tissue points on 20 frame pairs
STIRstereo endoscopic video, start to end of clip, porcine in and ex vivohundreds of clips, >3,000 pointsinfrared tattoo markersboth imagesdirect2025 test set CC BY 4.0; full set on IEEE DataPort, licence not statedopenwaits for a tool
video: waits for point tracking
Used in 2 of the 2D and microscopy registration papers read (Endo-TTAP 2025; S3-Tracker 2026).
STIR Challenge 2024/2025 test setsStereo surgical video, IR tattoo point trackingSTIRC2025 test clips (new); STIRC2024 on Zenodo 14803158Infrared-ink tattoo points at start and end of each clipboth imagesdirectCC BY 4.0openwaits for a tool
video: waits for point tracking
status change for STIR: previously held-out test sets now public
SUN-SEG (video polyp segmentation)colonoscopy video clips with polyps, consecutive frames1,013 clips, ~159K frames (from SUN database)polyp masks per frame plus boxes and scribblesboth imagesrequestSUN database terms: non-commercial research, request required (unverified)non-commercialneeds a request
the data needs a request or a data use agreement
Polyp masks on every frame give weak label truth for frame-to-frame alignmentCOLMAP on clips for sparse correspondences; two people click 8 points on 30 frame pairs
SuPer (surgical perception)stereo dVRK video of chicken tissue being grasped and tugged1 trial, 52 annotated frames20 tissue points tracked per annotated frameboth imagesdirectnone statedaskwaits for a tool
video: waits for point tracking
Tiny; access unverified. Licence checked on 2026-10-10 at https://www.sites.google.com/ucsd.edu/super-framework/dataset; derived test pairs shareable: unclear; commercial use: unclear. Google Drive link on the project site and in the Python-SuPer README, no terms; repo has no licence
SurgicalMotion point-tracking set (Tracking Everything in Robotic-Assisted Surgery)in vivo robotic surgery video, long-term point tracking20 videos x about 60 frames25 manual points per frame (15 tissue, 10 instrument) with occlusion flagsboth imagesdirectCC BY-NC-ND 4.0non-commercialwaits for a tool
video: waits for point tracking
VL-SurgPT (2025, 754 tissue clips, 17,171 points) also exists; its access is unverified.
SurgPose (articulated surgical tool pose with UV markers)da Vinci stereo video of 6 instrument types over ex vivo tissue; tool keypoint trackingAbout 120k instrument instances (80k train, 40k val), 1001-step trajectories7 keypoints per instrument from UV-fluorescent paint filmed under UV light on a replayed trajectory, SAM2-propagated and checked by peopleboth imagesdirectCC BY 4.0openwaits for a tool
video: waits for point tracking
Tracks instruments, not tissue; marker trick (replay under UV) mirrors STIR Licence checked on 2026-10-10 at https://doi.org/10.5281/zenodo.15278516; derived test pairs shareable: yes; commercial use: yes. Zenodo record linked from the SurgPose GitHub README; the CC BY-SA on surgpose.github.io covers only the website
SurgT (MICCAI 2022 soft-tissue tracking)stereo endoscopic video, soft-tissue tracking157 stereo videos from 20 casesbounding boxes of a tracked tissue region (one per frame)both imagesdirectnone stated for the data; benchmarking code Apache-2.0askwaits for a tool
video: waits for point tracking
Boxes only, weak for registration; access unverified. Used in 2 of the 2D and microscopy registration papers read (Endo-TTAP 2025; S3-Tracker 2026). Licence checked on 2026-10-10 at https://github.com/Cartucho/SurgT_benchmarking; derived test pairs shareable: unclear; commercial use: unclear. the benchmarking tool downloads videos and ground truth from Imperial College Box links with no data terms; challenge Data page returned 403; ask the organisers
Synthetic 2D speckle-tracking echo database (Alessandrini 2018, KU Leuven)simulated 2D echo sequences (apical 2/3/4-chamber), vendor-specific texture from 7 systems, healthy and ischemic motion105 sequences, one cardiac cycle eachground-truth motion from an electromechanical heart model (dense displacement / mesh points)both imagesdirectCC BY-SA 4.0openwaits for a tool
video: waits for point tracking
Exact known motion; link from paper (bit.ly/SyntheticDataQA2DSTE resolves to KU Leuven box); IEEE TUFFC 2018 doi 10.1109/TUFFC.2017.2786300 Licence checked on 2026-10-10 at https://doi.org/10.48804/XRJQK2; derived test pairs shareable: yes; commercial use: yes. KU Leuven RDR record by M. Alessandrini (105 sequences, 7 vendors, ground-truth tracking), DataCite rights; share-alike
Synthetic Deformable Tubular Benchmark (colonoscopy-like, FEM)synthetic endoscopy-like tubes with FEM deformation, moving and static cameras42 streams of 90 frames, 5,580 RGB imagesmaterial-point correspondences, optical flow, depth, poses, meshes (exact, synthetic)both imagesdirectCC BY 4.0openwaits for a tool
video: waits for point tracking
published 2026-10-10; generator code at icirauqui/LumenMorphalready exact; pair it with a real colonoscopy set for realism
TAS-1K simulated echocardiography (Sherbrooke 2026)SIMUS-simulated 2D echo sequences driven by meshes from real CAMUS and CARDINAL videos; known motion1,478 videos, 69,535 framesReference myocardial motion as temporal meshes, by constructionboth imageshiddenNot statedunknownleft out
its images or its truth are not released
Release promised on acceptance
TEyeD (near-eye video images)Head-mounted near-infrared eye videos, frame to frame, same person>20 million framesPupil, iris and eyelid landmarks, segmentations and gaze per frame (semi-automatic)both imagesdirectnone statedaskwaits for a tool
video: waits for point tracking
Eye-tracking, not retinal; landmarks on every frame could score frame-to-frame warps of the eye region. Low priority. Licence checked on 2026-10-10 at https://arxiv.org/abs/2102.02115; derived test pairs shareable: unclear; commercial use: unclear. paper gives Tuebingen and TUM download links; download folders hold no licence file; part of the images come from other eye datasets with their own terms
Thyroid Ultrasound Cine-clip (Stanford AIMI)2D thyroid ultrasound cine clips (probe sweep), 167 patients, frames of the same clip192 nodules, one cine clip eachradiologist nodule segmentations on the clip framesboth imagesaccountStanford research use agreement (terms not shown)non-commercialneeds an account
the download needs an account
Sweep frames are out-of-plane apart, so only nearby frames are 2D-registrable; label truth onlynodule masks give a weak check on frames a few apart; two people click 5 landmarks (vessels, cartilage) on 20 frame pairs
TUS-REC 2024/2025 (Trackerless freehand US reconstruction)freehand 2D ultrasound sweeps of forearms, same subject; estimate frame-to-frame rigid transformsabout 100 volunteers x 24 scans (training part 2 alone: 1,200 scans, 50 subjects); every frame has a poseoptical tracker (NDI Polaris) 4x4 pose per frame plus probe calibration, recorded by organisersboth imagesdirectCC BY-NC-SA 4.0 (Zenodo training records)non-commercialwaits for a tool
video: waits for point tracking
Pose estimation rather than image-to-image registration. TUS-REC2025 data not yet allowed for public use.
VFSSKep (videofluoroscopic swallowing keypoints)Videofluoroscopic swallowing study videos (30 fps), frames of the same swallow; hyoid, cervical vertebrae, soft palate288 videos; 13,895 labelled frames from 210 videos + 31,536 unlabelled framesPer-frame anatomical keypoints (hyoid, vertebrae, soft palate) annotated and checkedboth imageshiddenunverifiedunknownleft out
its images or its truth are not released
Code repo exists; paper says code and data will be made public (not yet confirmed). Vertebrae rigid, soft palate deformable: a good mixed-motion 2D case.
VL-SurgPT (in vivo surgical point tracking with ICG markers)in vivo da Vinci gastrointestinal surgery clips, tissue and instrument points tracked across frames754 tissue clips (180.5k frames, 1,862 tracks, 17,171 points) + 154 instrument clipsICG fluorescent markers at clip start and end (physical) plus two clinicians' manual tracks at 1 fps with status labelsboth imagesdirectnon-commercial (project page 'Licensing')non-commercialwaits for a tool
video: waits for point tracking
Start/end ICG points are marker truth like STIR; intermediate tracks are manual Licence checked on 2026-10-10 at https://szupc.github.io/VL-SurgPT/; derived test pairs shareable: unclear; commercial use: no. 'The original dataset and annotations of VL-SurgTPT cannot be used for commercial purposes'; no named licence, so sharing terms are unclear
XACV (DeNVeR coronary video set)Coronary X-ray angiography videos; frames of the same rununverified number of videosManually labelled vessel segmentation on video framesboth imagesdirectnone statedaskwaits for a tool
video: waits for point tracking
Host is a personal Google Drive link. Licence checked on 2026-10-10 at https://github.com/kirito878/DeNVeR; derived test pairs shareable: unclear; commercial use: unclear. README gives a Google Drive download; repo has no licenceBifurcation points clicked by two people on 20 frame pairs

Microscopy · 190

DatasetProblemPairsChecked againstAccessLicenceStatusNotesTo make the truth
10x Genomics Visium HD public datasetsVisium HD (2 um bins) with high-resolution H&E (or IF) of the same section and the CytAssist image; human and mouse tissuesMany datasets: microscope H&E vs CytAssist image vs bin gridSpace Ranger fiducial-based alignment (algorithmic, manual option in Loupe)one imagedirectCC BY 4.0 (10x dataset pages)opentruth to make
same section seen by the microscope and the CytAssist camera, near affine once aligned; Space Ranger's fiducial alignment is the provider's; worth it only with tissue points placed by people; files direct on cf.10xgenomics.com, CC BY 4.0
Fiducial frame on the slide gives physical markers Licence checked on 2026-10-10 at https://www.10xgenomics.com/datasets; derived test pairs shareable: yes; commercial use: yes. 10x dataset pages carry a CC BY 4.0 licence line (seen on several pages via search; 10x site returned HTTP 429 to direct fetches, so confirm on each chosen dataset page); cite per 10x guidelinesFiducial frame markers in both images give truth for the global transform; add 20 tissue landmarks for local error
10x Genomics Xenium public datasets with post-run H&EXenium in situ (DAPI/morphology images, transcript coordinates) and H&E of the same section afterwards; many human and mouse tissues (lung, colon, skin, lymph node, brain, Xenium 5K)Dozens of datasets, one same-section H&E-to-DAPI pair eachFor many datasets 10x ships an H&E-to-Xenium alignment matrix made with keypoints (unverified per dataset)one imagedirectCC BY 4.0 (10x dataset pages)opentruth to make
same-section DAPI vs post-Xenium H&E; 10x ships only a 3x3 similarity matrix (he_imagealignment.csv); nuclei pairs matched by a program and checked by a person would give a dense semi-automatic truth for a main board; files direct from cf.10xgenomics.com (checked on one dataset), CC BY 4.0, H&E OME-TIFF about 1.4 GB a dataset; STHELAR repackages 31 of them: build from one copy only
HEST-1k bundles many of them with alignments Licence checked on 2026-10-10 at https://www.10xgenomics.com/datasets; derived test pairs shareable: yes; commercial use: yes. same 10x dataset licence line as the Visium HD row; site returned HTTP 429, confirm on each chosen pageSame section: nuclei correspondences (DAPI vs hematoxylin) verified by a person on 10 regions; or apply known warp after alignment
10x Visium mouse brain serial sections (sagittal anterior/posterior)Visium with H&E, adult mouse brain, two serial sagittal sections each of anterior and posterior halves2 serial pairs (section 1 vs 2) plus anterior-posterior stitching pairsNone; Allen atlas used as visual reference in benchmarksone imagedirectCC BY 4.0 (10x dataset pages)openleft out
two serial pairs of one brain: too few pairs for a test set, no truth
Small and weakly annotated Used in 2 of the 2D and microscopy registration papers read (GPSA 2023; STalign 2023). Licence checked on 2026-10-10 at https://www.10xgenomics.com/resources/datasets/mouse-brain-serial-section-2-sagittal-anterior-1-standard; derived test pairs shareable: yes; commercial use: yes. 10x dataset licence line CC BY 4.0 seen on other 10x dataset pages via search; this page itself returned HTTP 429Map both sections to the Allen CCF with a public tool (e.g. DeepSlice) and compare region labels, a weaker check; or two people place 15 landmarks per pair
3D imaging mass cytometry serial sections (Kuett 2022 Nature Cancer)Imaging mass cytometry (40-plex) of 2 um consecutive sections, HER2+ breast carcinoma and lymph/blood vessel invasion models, for 3D reconstruction4 stacks: main model 152 sections (168 acquisitions), second 92 sections, plus two LVI models; ~240+ adjacent pairsAuthors' aligned stacks (Fiji Register Virtual Stack Slices, SIFT + similarity on SMA channel); no landmarksone imagedirectCC BY (Zenodo record)opentruth to make
4 serial IMC stacks with the authors' rigid alignment only (aligned and unaligned stacks released, CC BY, 6.7 GB); pairs at different depths of 4 specimens are weakly independent; two people would place vessel or duct points
Both unaligned and aligned stacks are released: rare for serial-section data; iridium (DNA) channel in every sectionTwo people place 20 landmarks (vessels, ducts) on 10 adjacent pairs; or use cell segmentation masks to score overlap of matched structures
3D-SpermFlagella (3D+t sperm flagellum centerlines)multifocal video microscopy of swimming human sperm, 3D flagellum centerlines over time135 tracked spermannotated 3D centerlines per time pointboth imagesdirectunverified (Zenodo)unknownwaits for a tool
3D+t flagellum centrelines of free-swimming cells: a tracking problem, unchanged; waits for point tracking
Sci Data 2026. Images are in the earlier 3D-SpermVid dataset.
3DeeCellTracker 3D time-lapse datasets (Wen 2021)3D+t images of moving/deforming worm brains, a beating zebrafish heart and a tumour spheroid; cell trackingConsecutive volumes in several sequencesManually corrected cell segmentations and tracksboth imagesdirectCC BYopentruth to make
3D+t worm-brain volumes on SSBD; manual labels only on the first volume, the later tracks are the software's output: a person checks tracked nuclei on chosen volume pairs
Also cites the Nguyen 2017 worm set on IEEE DataPort Licence checked on 2026-10-10 at http://ssbd.qbic.riken.jp/set/20190602/; derived test pairs shareable: yes; commercial use: yes. SSBD:repository record field 'License: CC BY license' (version not given)
ACROBATIHC to H&E breast whole-slide images, same tissue750 training, 100 validation, 303 test37,000+ landmarks, used only by the challenge's evaluationboth imageshiddenCC BY 4.0openleft outGigapixel 2D: could also serve out-of-core tests without landmarks. status change: Round 4 count: listed in 3 of the 38 round-4 surveys and benchmark papers (Chen 2025 survey (MedIA); WSI registration review 2025; ACROBAT 2023). Used in 5 of the 2D and microscopy registration papers read (ACROBAT 2022 challenge (8 methods: Gestalt, VALIS, AGHSSO, MEVIS, NEMESIS, MeDAL, SK, MFRGNK); VALIS 2023; DeeperHistReg/RegWSI 2024; STAR 2025; ...).
AIM SMLM drift correction data (Sci Adv 2024)SMLM localisation data (DNA origami PAINT, 3D microtubules, colon tissue, cell nuclei) and a simulated set, drift over acquisition4 experimental + 1 simulated datasetsimulated: known drift; DNA origami geometry as structural checkboth imagesdirectCC0openwaits for a tool
localisation lists, not images: unchanged; waits for mesh/point-cloud scoring
Also on Dryad; MATLAB .mat localisationsorigami grid spacing gives an internal ruler for residual drift
Allen Brain Observatory two-photon (Visual Coding / Visual Behavior)in vivo two-photon calcium movies of mouse visual cortex, frame-to-frame motion correctionhundreds of sessions (each ~100k frames)per-frame rigid x,y translations from the Allen pipeline (method output, not independent truth); ROI masksone imagedirectAllen Institute terms of use (non-commercial research; unverified for raw movies)non-commercialwaits for a tool
two-photon movies whose only motion reference is the pipeline's own offsets; non-commercial terms: unchanged; waits for point tracking
Offsets are a strong baseline to compare against, not truth; raw movies needed for a benchmarkinject known non-rigid motion into motion-corrected movies
Allen whole mouse brain MERFISH with CCF registration (MERFISH-C57BL6J-638850, Yao 2023)MERFISH coronal sections covering one whole mouse brain, each section registered to Allen CCFv3 (global affine plus section-wise deformable); Zhuang ABCA-1 to 4 brains also available~59 sections to atlas, plus consecutive-section pairs; Zhuang brains give between-animal pairsAuthors' section-to-CCF mappings (point-to-point) and CCF region labels per cell; manually set midlineboth imagesdirectCC BY-NC 4.0non-commercialleft out
same Yao 2023 MERFISH-to-CCF data as the 'ABC Atlas MERFISH whole mouse brain sections mapped to CCFv3 (Allen, Yao 2023)' row
Truth is an algorithmic registration, not expert landmarksUse authors' mappings as reference and verify with 20 expert landmarks on 10 sections
Angelo-lab MIBI-TOF cohorts with serial H&E (Risom 2022 DCIS, Keren 2018 TNBC, McCaffrey 2022 TB)MIBI-TOF 36-37 plex images of TMA cores or FOVs; H&E on serial sections used by pathologists to pick regionsDCIS: about 79 patients' cores; TNBC: 41 patients; TB: 13 patients x 2 FOVsnone for registration; pathologist region masks on MIBI (DCIS)one imagedirectMendeley Data (CC BY 4.0, unverified); TNBC via mibi-shareopenleft out
the linked DCIS record holds MIBI images, masks and tables but no H&E, so no second image
released files are MIBI images and masks; serial H&E images do not appear to be released (unverified), so mostly fails the pair criterionif H&E released: duct or granuloma outlines on both by a person; otherwise none
AOSLO cone photoreceptor images with cone locations (Roorda lab)adaptive optics scanning light ophthalmoscope images of the cone mosaicfew images (count unverified)x,y location of every cone per imagedirectCC BY 4.0openleft out
4 single AOSLO images of different eyes, no second image of the same patch
If repeated images of the same patch exist, cone positions are exact correspondences; otherwise only single imagesknown eye-motion warp (strip-wise shifts) of a real AOSLO image with known cone positions
Articular cartilage resonance-scanned microscopy deformation validation (lsmgridtrack)laser scanning microscopy of cartilage and cells under dynamic deformation, frame to framevalidation tests (count unverified)validation outputs of the grid tracking tool (type unverified, likely known applied deformations)direct: Zenodo, one tar.gz of 8.3 GB (not readable by ranges), no accountCC BY 4.0 (Zenodo)openleft out
The record is one 8.3 GB tar.gz of inputs, scripts and outputs of lsmgridtrack 0.3's validation, and the tool's test (test_random_resonance.py) warps a single undeformed, grid-photobleached cartilage image (Zenodo 6423385, 90 MB) by random B-splines with a fixed public seed. One specimen and a public deformation generator: nothing the corpus could not draw itself on one image.
inputs, scripts and outputs of lsmgridtrack v0.3photobleached grid lines, if present, are physical markers; else a known warp
ARTISTA axolotl telencephalon Stereo-seq atlas (Wei 2022)Stereo-seq spatial transcriptomes of axolotl telencephalon: regeneration time course (2-60 days post injury) and development stages; serial sections, same and different animals3-4 serial sections per time point (2, 5, 10, 15, 20 DPI) plus 30 and 60 DPI and stages 44, 54, 57, juvenile, adultSpatial clusters and cell-type annotations per section by the authors; no point correspondencesboth imagesdirectnone stated on STOmicsDB pagesasktruth to make
cluster labels per section, no points; whether the ssDNA images are downloadable is unverified and STOmicsDB names no licence (ask)
Download page and raw data on STOmicsDB; whether ssDNA images are released is unverified Licence checked on 2026-10-10 at https://db.cngb.org/stomics/artista/; derived test pairs shareable: unclear; commercial use: unclear. ARTISTA download pages name no licence; CNGBdb terms and data sharing policy are JavaScript-only pages and could not be read; ask BGI/CNGBTwo people place 10 landmarks (ventricle tips, pallial boundaries) on adjacent sections; label overlap after alignment as a weaker check
ASHLAR / MCMICRO exemplar CyCIF data (normal colon, tonsil)multi-cycle cyclic immunofluorescence whole-slide tiles: tile stitching plus cycle-to-cycle registration via repeated Hoechst channelcolon: 9 cycles x 609 tiles (61 GB); tonsil and EMIT exemplars on Synapserepeated nuclear stain in every cycle (same cells) lets nuclei centroids act as correspondences; stage positions in metadataboth imagesaccountnone statedaskneeds an account
the download needs an account
Multimodal-by-construction: every cycle repeats Hoechst, giving exact same-cell truth up to segmentation status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (WSI registration review 2025). Licence checked on 2026-10-10 at https://doi.org/10.7303/syn25826362; derived test pairs shareable: unclear; commercial use: unclear. Synapse folder (project 'Ashlar') has no access requirement, no licence annotation and no wiki; download needs a free Synapse account; ASHLAR code MITsegment nuclei in Hoechst of cycles 1 and N, match by Hungarian assignment, check by eye on 200 nuclei
ASHLAR colon CyCIF raw tiles (Muhlich 2022)9-cycle t-CyCIF of normal human colon (28-plex), RareCyte scanner, Hoechst every cycle; also used for tile stitching9 cycles x 609 tiles (5481 tiles); cycle pairs and tile-overlap pairsNone; authors measured residual error by block-wise optical flow on the Hoechst channel (median 0.12 um for ASHLAR)one imageaccountnone statedaskneeds an account
the download needs an account
Synapse download needs a free account; the paper also re-scanned one H&E slide twice on 4 scanners (location of those scans unverified) Licence checked on 2026-10-10 at https://doi.org/10.7303/syn25826362; derived test pairs shareable: unclear; commercial use: unclear. same Synapse folder as the ASHLAR/MCMICRO exemplar row; no data licence; the article is CC BY 4.0Nuclear centroid matching across cycles, verified by a person on 10 tiles; overlap regions give tile-to-tile truth by stage coordinates
ATOM (histology-to-organ benchmark)Histology section location within organ image volume, multi-institutionUnverifiedKnown source location of subregionsboth imagesdirectcode MIT; data (Google Drive data_pkl) no licence, built from MedMNIST and ADNIaskwaits for a tool
2D planes to place back in 3D volumes; the ADNI part needs a data use agreement: unchanged; waits for 2D/3D pose error
status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (ATOM 2024). Licence checked on 2026-10-10 at https://github.com/haizailache999/Image-Registration; derived test pairs shareable: unclear; commercial use: unclear. README sends users to MedMNIST and ADNI; ADNI's data use agreement forbids redistribution, so derived pairs likely cannot be shared
AURORA-Metastatic-Breast-Multiomics (TCIA 2025)H&E slides of primary and metastatic breast tissue, some samples with 2-3 images, plus HLA-A immunofluorescence on 37 samplesabout 29 samples with repeat H&E and 37 H&E-IF combinationsnoneone imagedirectunverified (TCIA collection licence)unknowntruth to make
repeat H&E and H&E-IF images whose section relation is unclear, no truth, TCIA licence unverified
Weak: unclear whether repeat images are the same section.two people place 10 landmarks per H&E-IF pair
Automated EM-LM correlation with fiducial particlesImmunofluorescence vs EM tiles of same HEK293 sectionsUnverified countFiducial particle coordinates; manual ec-CLEM correlationsboth imagesBioImage Archive, direct HTTP or FTP per fileCC BY 4.0 (BioImage Archive study)opento build apartpriority 3
One grid of 90 nm HEK293 cryosections: four positions, each an immunofluorescence tile (1 to 1.6 MB) and an EM tileset (bin4 0.1 to 0.3 GB, full up to 3.5 GB), with ec-CLEM similarity transforms for positions 1 to 3 and EM fiducial-particle points for positions 1 and 3 (BioImage Archive file lists). Four pairs of one grid, scored on a similarity fit: far under ten pairs.
AxonEMserial-section EM of mouse and human cortexnone between volumesaxon instance masksone imageaccountMIT (Hugging Face dataset cards pytc/AxonEM and pytc/EM30)openneeds an account
the download needs an account
Segmentation set; listed by the point-cloud survey. Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Point-cloud shape survey 2025). Licence checked on 2026-10-10 at https://huggingface.co/datasets/pytc/AxonEM; derived test pairs shareable: yes; commercial use: yes. challenge page links the images and training labels on Hugging Face, both cards declare license: mit; human volume comes from surgical tissue, check the paper for consent limitsIsotropic enough to cut virtual sections and apply known per-section warps: exact truth for serial-section alignment.
BCI (Breast Cancer Immunohistochemical image generation)Consecutive sections, H&E vs HER2 IHC, breast cancer4870-4873 registered 1024x1024 patch pairs (3896 train, 977 test)Patches pre-registered by the authors; structural-level alignment only (adjacent sections)one imagedirectBCI dataset licence: free for non-commercial purposesnon-commercialtruth to make
adjacent-section H&E/HER2 patches pre-aligned by the authors with documented residual misalignment (a known warp would inherit it); whole slides not released; non-commercial
Residual misalignment from consecutive sections is documented (HEMIT paper); WSIs not released Licence checked on 2026-10-10 at https://bupt-ai-cz.github.io/BCI/; derived test pairs shareable: unclear; commercial use: no. dataset page and download_dataset.md: 'freely available to academic and non-academic entities for non-commercial purposes such as academic research'; code BSD-3-ClauseTwo people place 10 matched landmarks per patch on a 50-pair test subset; or nuclei detection on both and verified matching
Benchmark datasets for 3D MALDI and DESI imaging MS (Oetjen 2015, MTBLS176)3D MSI from serial sections: mouse pancreas, mouse kidney, human oral squamous cell carcinoma (MALDI), human colorectal adenocarcinoma (DESI) with H&E of the same sections3 tissue MALDI series (mouse kidney, mouse pancreas, human OSCC; tens of sections each, exact counts unverified) and 26 DESI sectionsDESI HDF5 holds the authors' co-registered MS and H&E images; MALDI sets have 3D voxel positions from their own registrationone imagedirectGigaDB CC0 (unverified)opentruth to make
serial MSI sections and same-section H&E with the authors' co-registration only; MSI images must be rendered from spectra; ablation pattern or vessel points to be placed by people
section-to-section and MSI-to-H&E registration; also in MetaboLights MTBLS176H&E after MSI on the same section: ablation pattern and tissue edges give near-exact MSI-H&E points; serial sections: vessel landmarks by two people
Benchmark light microscopy matching TEMBlock, section and ROI fluorescence vs TEM, mouse brainUnverifiedNone stateddirectCC0openleft out
one block (experiment 373, 7 light-microscopy files); the TEM images are not in the record, and block to section is a 2D/3D placement
Truth unverifiedTwo people click 10 cell or vessel points on fluorescence and TEM of 10 regions (inter-observer error reported).
BIFROST / Functional Drosophila AtlasIn vivo multiphoton fly brain to atlas/JRC2018, inter-subjectReplication dataset + atlasGenetically labelled cell types used as ground truthboth imagesdirectCC0openleft out
below the truth bar: its 9 released flies make 4 pairs, and the truth is the two LC11 glomerulus centres a case; more animals would be needed, not a better truth
BigStitcher 5-angle SPIM exampleMultiview light-sheet of flatworm, 5 angles5 viewsFluorescent beads (561 channel)both imagesdirectCC BY 4.0opentruth to make
bead positions are not given (a bead detector and matcher is needed); one specimen, five views
Bead-based multiview registration
BigStitcher-Spark example datasets3D tile stitching and interest-point registration of light-sheet tilesStitching set: 18 tiles of 512x512x86; interest-point set (unverified size)Reference BigStitcher solution (aligned XML) at each stage; not independent truthone imagedirectcode BSD-2-Clause; example data (Google Drive) no licence statedasktruth to make
small 3D light-sheet tiles, only BigStitcher's own solution as reference: known shifts or warps on the tiles would make truth
Unaligned vs aligned XML give stage-like initial and solved positions. Licence checked on 2026-10-10 at https://github.com/PreibischLab/BigStitcher-Spark; derived test pairs shareable: unclear; commercial use: unclear. README 'Example Datasets' section gives Drive links only; the BSD licence reads as a code licenceApply known translations/smooth warps to tiles and score recovered positions; or compare to the aligned XML as a weak reference.
Breast cancer serial sections for 3D reconstructionSequential WSI of breast tumour~20 sectionsNone statedone imagedirectCC BY 4.0opentruth to make
about 20 serial sections of one tumour, CC BY 4.0, 5 GB, no truth; one specimen, so pairs are weakly independent
Two people place 15 landmarks (vessels, duct and gland points) on 10 consecutive section pairs.
CaImAn benchmark datasets with manual annotationsin vivo two-photon (and one-photon) calcium imaging movies of mouse brain, frames with motion9 annotated two-photon datasets plus othersconsensus neuron footprints from multiple labelers (no motion truth)one imagedirectCC BY 4.0openwaits for a tool
calcium movies without motion truth: unchanged; waits for point tracking
Motion truth absent; neuron masks are on the motion-corrected mean, a weak checkadd known rigid and piecewise-rigid motion to a stabilised movie; or NAOMi simulation
CAST spatial omics reproduction data (STARmap PLUS, MERFISH, Slide-seq mouse brain)Single-cell spatial transcriptomics of mouse brain sections across platforms and animals (STARmap PLUS, MERFISH, Slide-seq), for cross-sample alignmentMany section pairs within and across platforms (counts unverified)Cell-type and region labels; authors' CAST alignmentsboth imagesdirectCC BY 4.0openwaits for a tool
cell tables and AnnData, no images: unchanged; waits for mesh/point-cloud scoring
Point clouds and cell tables rather than images (unverified)Region labels on both; verify with 20 expert landmarks on 10 pairs
Cell Tracking Challengetime-lapse cell microscopymanycells move independently: no deformation truthone imagedirectCTC conditions of use of the imagesaskleft out
a tracking benchmark: cells move independently, so no deformation to register; its terms also forbid copying the data or annotations
Licence checked on 2026-10-10 at https://celltrackingchallenge.net/datasets/; derived test pairs shareable: no; commercial use: no. only CTC-related use is free; public non-CTC scientific use needs organiser permission, commercial use needs organisers and data providers; 'Cloning of our datasets or their parts, including reference annotations, is strictly forbidden'Cells move independently: no deformation truth; skip.
Cervical Cancer Cell Image Database multi-focus stacks (CCCID)Pap-stained cervical cytology cell images, 11-layer focus stacks (1 um step)single cells in 4 Bethesda classes (count unverified)nonedirectunverifiedunknownleft out
11 focal planes of one scan per cell, no second image to register
fails the pair criterion: focal planes of one scan, no deformation; listed for completenessnot worth making; at most a known warp between focal planes
CGCI H&E and IHC whole-slide collections in IDC (BLGSP, HTMCP)H&E and IHC (BCL2, BCL6, CD10, Ki-67, CD20, CD3, p16, others) slides of the same subjects: Burkitt lymphoma (388 subjects, 1933 slides), cervical cancer (211 subjects, 525 slides), plus DLBCL and lung subsetsMany H&E-IHC slide pairs per subject; serial-section status unknownNoneone imagedirectCC BY 4.0opentruth to make
H&E and IHC slides of the same subjects but block pairing unknown; same-block pairs must be picked by eye before two people place points
Weak: slides may come from different blocks; FFPE and frozen mixed. Only pairs from the same block are registrable Licence checked on 2026-10-10 at https://doi.org/10.5281/zenodo.17381396; derived test pairs shareable: yes; commercial use: yes. Zenodo record for the IDC DICOM conversion of CGCI-BLGSP slides (rights CC BY 4.0); HTMCP collection not checked separatelySelect pairs from the same block by visual check, then two people place 20 landmarks on 10 pairs
CLEM-Regvolume EM to fluorescence, HeLa cells3expert landmark pairs (145, 42, 55)both imagesEMPIAR and BioStudies, direct HTTP, FTP or AsperaEM: CC0 (EMPIAR policy); light: CC0 for S-BSST1175, none stated for S-BSST707 and S-BSST1075askin the corpuspriority 2
Three HeLa cells imaged by Airyscan then FIB-SEM at 5 nm (EMPIAR-10819, EMPIAR-11537) or SBF-SEM (EMPIAR-11666), with manual BigWarp landmark pairs (145, 42, 55) deposited as CSV next to the light data (S-BSST707, S-BSST1075, S-BSST1175) and, for EMPIAR-10819, EM stacks downscaled to 10 and 20 nm by the provider (Nature Methods 2025, EMPIAR and BioStudies entries). The only volume-EM to light pairs with an expert truth, but three cells; EMPIAR data are CC0 by policy and S-BSST1175 says CC0, while S-BSST707 and S-BSST1075 state no licence.
Licence checked on 2026-10-10 at https://www.ebi.ac.uk/empiar/policies/; derived test pairs shareable: yes; commercial use: yes. EMPIAR-10819 record shows no per-entry licence; EMPIAR policy: 'All data in EMPIAR is freely and publicly available ... under the CC0 licence'
CODEX imaging of HCC (TCIA)37-plex CODEX whole-slide images of 15 fresh-frozen hepatocellular carcinomas plus whole-slide images (H&E)15 patients (plus spleen and lymph node controls)noneone imagedirectunverified (TCIA, likely CC BY 4.0)unknowntruth to make
875 GB of CODEX and H&E whose section relation is unclear, no truth, TCIA licence unverified
page lists whole slide images alongside CODEX; same or adjacent section unverifiedtumour-liver interface traced on both by a person; nuclei matching if same section
Colorectal cancer TMA IHC/ISH/H&E epithelium set (Pettersen, DataverseNO)Tissue microarray cores (serial TMA sections) stained H&E, 13 IHC markers and 4 ISH probes, colorectal normal and cancer, same patient across stains100 patients x 6 cores x up to 18 stains (13,179 core images); any two stains of the same core form a pairEpithelium segmentation mask on every core image (pathologist-annotated per the record)both imagesdirect: DataverseNO API (S3 redirect, range requests), one zip per stain (16 to 90 GB), no accountCC0 1.0 (DataverseNO and README)openin the corpuspriority 2
built as ColorectalTMAStains (track colorectal-tma-automatic); row status was never updated
Labels on both images allow Dice after warping; cores from different TMA sections, so non-rigid differences and occasional missing coresLabels already on both images; add 10 landmarks per core on a 50-core subset for a TRE
Confocal and FIB-SEM CLEM sample (Weiner, S-BSST456 with EMPIAR-10479)confocal volume vs FIB-SEM volume of the same cells1 sample pairnone; the book chapter shows a manual Amira correlationdirectunverifiedunknowntruth to make
one 3D confocal and FIB-SEM pair of the same cells (50 MB and 2.8 GB), no correspondences: two people mark 10 organelle landmarks in both
tutorial data for MCB volume 162 (CLEM IV)two people mark 10 organelle landmarks in both volumes
CORE multi-IHC colorectal cancer cohort (UHCW)Colorectal cancer slides stained H&E, CK8/18 and mismatch-repair IHC (MLH1, MSH2, MSH6, PMS2); serial sections, same block86 patients x 16 slides; 7 cases x 6 slides usedNot statedhiddenNot statedunknownleft out
its images or its truth are not released
Private (University Hospitals Coventry and Warwickshire)Nuclei centres detected by a program on both slides then verified by a person at 20 points per pair
Coregistered histology and DTI of meningioma (Lund)ex vivo 200 µm DTI and histology sections of meningioma, same tumourtumour count unverifiedhistology co-registered to MRI by the authors (method unverified)directcontrolled access (AIDA data sharing agreement); AIDA BY licence within AIDAaskneeds a request
data at the AIDA data hub under controlled access (email and signed data sharing agreement); the pairs are 2D (H&E vs one DTI slice) with landmarks
Data on GitHub (jan-brabec/microimaging_histology_DIB). Licence checked on 2026-10-10 at https://datahub.aida.scilifelab.se/10.23698/aida/micromen; derived test pairs shareable: unclear; commercial use: unclear. AIDA page: 'Controlled access. Free for use in legal and ethical medical diagnostics research', access by email and signed data sharing agreement; CC BY-NC-SA 4.0 applies to the processing code onlytwo people mark 10 matching points per section
Correlative 3D X-ray histology of placentas (S-BIAD1270, S-BIAD2433)micro-CT (3D X-ray histology) vs histology and electron microscopy of the same placenta, mouse to rhinocount unverifiednone stateddirectCC BY 4.0openwaits for a tool
2D sections against a 3D micro-CT of the same placenta: unchanged; waits for 2D/3D pose error
multiscale correlative workflowvessel bifurcations in micro-CT and the matching section by two people
Correlative cryo-SIM and soft X-ray tomography (S-BIAD18-20, S-BSST689, EMPIAR-10839)cryo 3D-SIM fluorescence vs cryo soft X-ray tomography of the same cell (reovirus, F-actin, NIH-3T3)a few cells per entrycorrelated volumes given (EMPIAR-10839 stitched and correlated); fiducial-based, unverifieddirectCC0 (EMPIAR default, unverified)opentruth to make
3D cryo-SIM (S-BSST728) and soft X-ray tomogram (EMPIAR-10839) of one cell, only the authors' correlated volume: points picked in both and checked by a person
Diamond B24 correlative workflow; S-BIAD18-20 and S-BSST689 hold related reovirus and actin studiesfiducial beads visible in both modalities, picked by a program and checked by a person
CosMx SMI public FFPE datasets (NanoString/Bruker)CosMx single-cell spatial transcriptomics/proteomics with morphology fluorescence images; human colon, breast, lymph node, liver, tonsil, brain, kidney, pancreas, skin, NSCLC; mouse brain (2 coronal sections)Few: mouse brain 2 sections; H&E of the same or adjacent section not statedNone for registrationone imagedirectno dataset licence; site 'All rights reserved' and website terms of useaskleft out
no H&E of the same section is stated, only two mouse brain sections, and the site reserves all rights
Weak or failing for H&E registration: page does not say H&E is included; download route not described on the page Licence checked on 2026-10-10 at https://www.brukerspatialbiology.com/products/cosmx-spatial-molecular-imager/ffpe-dataset/; derived test pairs shareable: unclear; commercial use: unclear. datasets page offers downloads with no data termsIf an adjacent H&E exists, two people place 20 landmarks on 10 FOVs; otherwise only morphology-channel-to-transcript-density checks
CRC_FFPE-CODEX_CellNeighs (Schurch 2020, TCIA)56-marker CODEX of colorectal cancer TMAs (140 regions, 35 patients) with histopathology images200 images, 35 patients; H&E of the TMA listed as histopathologynoneone imagedirectCC BY 4.0opentruth to make
CODEX TMA hyperstacks with H&E whose section relation is unverified, 2 TB, no truth
processed CODEX hyperstacks (cycles already registered); whether H&E is the same section is unverified; Aspera downloadcore-level H&E to CODEX DAPI: nuclei matching, person verifies 20 per core; core outlines give a coarse check
CREMI (A, B, C and A+, B+, C+)serial section TEM volumes, section-to-section alignment (padded raw has more context; realigned versions from FlyWire)6 volumes (A, B, C and A+, B+, C+), 125 sections eachneuron IDs, synaptic clefts and partners on every section (proof-read)both imagesdirectnone statedaskin the corpuspriority 3
In the corpus: 10 test and 10 development sections of the unlabelled test volumes A+, B+ and C+, each paired with a known in-plane warp of itself (as SNEMI3D); ranked apart in serial-em-synthetic. The labelled volumes A, B and C are CREMIClefts. No licence stated (ask).
Neuron labels continue across sections, so label overlap after alignment is a dense check; known misalignments in the original Licence checked on 2026-10-10 at https://cremi.org/data/; derived test pairs shareable: unclear; commercial use: unclear. data page lists HDF5 downloads and acknowledgements (FAFB imaging by Zheng, Bock et al.) with no termsknown in-plane warp of each section of the realigned volume gives exact truth
CRLM mid-infrared, MALDI lipid MSI and adjacent H&E (Rittel 2026, CeMOS)FT-IR (mid-infrared) imaging and MALDI MSI on the same fresh-frozen section, H&E on adjacent section; colorectal liver metastases12 specimensnone releasedone imagedirectCC BY 4.0opentruth to make
12 specimens of IR and MALDI on one section with adjacent H&E, CC BY 4.0, 21.6 GB, no truth; spectral images to render
three modalities: IR-MSI same section, H&E adjacenttumour-liver boundary traced on all three; vessel landmarks by two people
Cryo soft X-ray tomography with TetraSpeck fiducials (EMPIAR-10615, 10624)cryo soft X-ray tomograms of NIH-3T3 cells with TetraSpeck beads (and 100 nm gold) as correlative fiducials for cryo fluorescence (3D-SIM)a few tomograms of cells and of beads on grids, across 2 entriesnone released; beads visible in X-ray and fluorescencedirectEMPIAR (CC0 by archive policy, unverified for these entries)opentruth to make
bead positions are not released and would have to be detected and matched across X-ray and fluorescence; a few tomograms
Whether the matching fluorescence (cryo-SIM) volumes are deposited is unverified; without them it fails the pair criterion.bead centroids in SXT and fluorescence volumes = landmarks (beads are fluorescent and X-ray dense); fit an affine on half, test on the rest
Cryo-4i frozen mouse intestine multi-round images (SSBD 491)5 rounds of iterative indirect immunofluorescence on one frozen section; round-to-round registrationrounds per field of view (count unverified)none; DAPI in every roundone imagedirectCC BY 4.0opentruth to make
3D confocal z-stacks (100 planes) of one frozen section over 5 staining rounds with a nuclear stain each round, no truth: nuclei matched across rounds and checked by a person
cycle registration on the repeated DAPI channel is the classic multiplex alignment tasknuclei centroids from DAPI in each round matched automatically, person checks a subset
DBiTplus same-section spatial omics with CODEX and H&ESame tissue section: DBiT-seq spatial transcriptomics, then multiplexed IF (PhenoCycler/CODEX) and H&E; mouse embryo, mouse cerebellum, human lymph node~3-4 sections, each H&E vs mIF (and vs DBiT spot grid)None released; authors registered images to the DBiT gridone imagedirectCC BY 4.0openleft out
three or four sections of three tissues: too few pairs for a test set
Same section, so near-exact correspondence; microfluidic channel pattern also visibleSame section: nuclei in DAPI vs hematoxylin give dense truth after alignment; the DBiT channel grid acts as fiducials for spot-to-image truth
Deep Learning Virtual Staining Dataset: breast, mouse tumour, organoids (Ye and Sun 2025)autofluorescence or multi-channel fluorescence and H&E of the same samples; human breast, mouse tumours, organoidsunverified count; raw .mrxs slides plus processed patchesprocessed patch pairs aligned by the authorsboth imagesdirect: Zenodo record 17646116; zips readable by ranges (one AF/H&E slide pair about 1.5 GB)CC BY 4.0 (Zenodo)opentruth to makepriority 3
ten unaligned whole-slide pairs with no points and nothing aligned by construction: people must place points
raw whole slides (Initial_data.zip) allow real registration testssame section: landmarks on 10 slide pairs by two people, or known warp on the aligned patches
DeepCLEM training dataFluorescence vs SEM of C. elegans sections121 manually aligned pairsManual alignmentboth imagesdirect: Zenodo record 6973994, one 17 MB zipCC BY 4.0 (Zenodo record and the zip's Readme)openin the corpuspriority 3
Zenodo (CC BY 4.0, one 17 MB zip): 117 training and 4 validation pairs of 256x256 tiles, Hoechst fluorescence and SEM of the same C. elegans sections, superimposed by hand in Inkscape (scaled, rotated and moved) by the authors (F1000Research 2020). Aligned by construction, the truth would be a known warp (synthetic); the file names point to two or three section series (A1, A2, s1 to s12), so ten test and ten development pairs cannot come from different animals.
DeepLIIF IHC and mpIF same-slide datasetSame slide: Ki67 IHC brightfield and multiplex IF (DAPI, LAP2beta, Ki67), bladder and lung tissue; plus breast cancer setCo-registered 512x512 tiles (hundreds of tiles; several hundred train/val/test)Tiles co-registered by the authors; segmentation masksone imagedirect: Zenodo record 4751737; zips readable by rangesCC BY 4.0 on the Zenodo deposit; the code's Apache 2.0 with Commons Clause (non-commercial) covers the code onlyopenin the corpuspriority 3
Zenodo (data CC BY 4.0; the DeepLIIF code is Apache 2.0 with Commons Clause): 512x512 tile sets of Ki67 IHC with hematoxylin, DAPI, LAP2beta and Ki67 mpIF of the same slides, the mpIF scaled and co-registered onto the IHC with affine transforms by the authors (575 training, 91 validation and 598 test tiles in the files), plus a breast set. The tile names (Lung_x_y, Bladder_x_y) point to few slides and the paper gives no patient count, so subjects are too few for ten plus ten; aligned by construction, the truth would be a known warp (synthetic).
Already-aligned tiles onlyAligned by construction: apply a known smooth deformation to obtain exact multimodal truth (IHC hematoxylin vs DAPI)
Dentin paired high- and low-resolution confocal (Data in Brief 2026)confocal fluorescence of the same tooth dentin fields at 100, 200, 400 and 800 nm pixel size6 regions x 18 slices at 4 resolutionsauthors registered the resolutions (not independent)one imagedirectunverifiedunknownleft outpriority 3
one tooth slice of one donor (six regions): one subject, and the pairs differ only by pixel size
Cross-resolution 2D pairs.known warps on the registered pairs; tubule centres matched by a program and checked by a person
Destain-restain cytology cell blocks, H&E to IHC (Al-Attar 2026, CytoJournal)FNA cell block sections scanned in H&E, destained, restained with one IHC marker, rescanned; same section9 patients, several stains eachmanual synchronisation in the viewer onlyone imagerequestavailable from the corresponding author on reasonable requestaskneeds a request
the data needs a request or a data use agreement
the paper calls it perfect pixel-level registration; sparse cells make it hardsame section: nuclei matching H&E vs IHC haematoxylin counterstain, person verifies 20 per slide
Developing human heart spatial transcriptomics (Asp 2019 Cell)Spatial transcriptomics with H&E, developing human heart at 3 developmental stages, serial sectionsSerial-section pairs per stage (counts unverified)Spot coordinates and transformation matrices to H&E images; no landmarksone imagedirectCC BY 4.0opentruth to make
serial ST/H&E sections of three stages with spot-to-image matrices only, CC BY 4.0; two people would place points on adjacent H&E pairs
Older ST (100 um spots) platform Licence checked on 2026-10-10 at https://data.mendeley.com/datasets/dgnysc3zn5/1; derived test pairs shareable: yes; commercial use: yes. Mendeley Data record (DataCite rights CC BY 4.0)Two people place 15 landmarks on adjacent H&E pairs
Diffractive-optics super-resolution microscopy pairswidefield and structured-illumination super-resolution images of the same cells, multispectral and hyperspectralfive records (counts unverified)pairs from the same acquisitionboth imagesdirectunverifiedunknownleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): its licence is not shown to allow commercial use (unverified). It was proposed as: widefield and SIM of the same cells, aligned by acquisition: a known warp only; licence unverified
Sci Data 2026.known warps on aligned pairs for exact truth
DL-SMLM (paired low-resolution and SMLM super-resolution)widefield frame stacks and SMLM super-resolution reconstructions of the same cells, six subcellular structures188 setspairs from the same acquisitionboth imagesdirect: figshare ndownloader, one zip per structure, readable by ranges, no accountCC BY 4.0 (figshare)openleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): open, but not among the six minor sets kept. It was proposed as: Per cell, a 256 x 256 TIRF widefield image (130 nm pixels, sum of a 100-frame stack also given) and the SMLM reconstruction of the same field at 8x (2048 x 2048, 16.25 nm, float32), plus the localisations, for six structures (CCPs, ER sec61b and KDEL, microtubules, mitochondrial outer and inner membranes); 188 cells in all (26 to 41 a structure). Both come from one camera and acquisition, so there is no registration to solve unless the corpus warps one side (synthetic). CC BY 4.0 on figshare, six zips (18.5 GB) readable by ranges.
Sci Data 2025.known warps on aligned pairs for exact cross-resolution truth
DLPFC Visium serial sections (spatialLIBD, Maynard 2021)10x Visium spatial transcriptomics with H&E, human dorsolateral prefrontal cortex, consecutive sections of the same donor12 sections from 3 donors (4 per donor): adjacent pairs 10 um apart and pairs 300 um apartManual cortical layer labels (L1-L6, white matter) on every spot of every section, by the authorsboth imagesdirectcode Artistic-2.0; raw images and counts (AWS, Globus) no licence statedaskin the corpus
Same data as the row spatialLIBD DLPFC Visium (Maynard 2021): in the corpus as DLPFCVisium, the layers rasterised from the spots.
Truth is layer labels, not point landmarks; full-resolution H&E images available with spot coordinates Licence checked on 2026-10-10 at https://github.com/LieberInstitute/HumanPilot; derived test pairs shareable: unclear; commercial use: unclear. HumanPilot and spatialLIBD repos are Artistic-2.0 (software); the AWS/Globus raw data pages give no termsLayer labels already on both; add 10 H&E landmarks (vessels) per adjacent pair for a TRE
Dual-modality IF and IMC whole-slide data with adjacent H&E (Kim 2023 Cell Rep Methods)whole-slide IF (DAPI, E-cadherin, CD45) and 36-plex IMC FOVs of the same slide, plus adjacent H&E; oesophagus and mouse brain (unverified)unverified; many IMC FOVs per slidenone independent; authors' cross-correlation then feature registrationone imagedirectCC BY 4.0opentruth to make
same-slide IF and IMC with adjacent H&E, the authors' registration only (CC BY 4.0, 10.8 GB); nuclei IMC DNA vs IF DAPI to match and check
IMC FOV to IF whole-slide placement is a 2D search-and-align tasknuclei matching IMC DNA vs IF DAPI on each FOV, person verifies
E-Staining DermaRepo (unstained and H&E skin biopsies)brightfield images of unstained skin biopsy sections and the same sections after chemical H&E87 whole-slide pairs, 22 patientsnone (file name pairing only)one imagedirectpublic for research; commercial use restricted (data article)non-commercialwaits for a tool
87 same-section unstained/H&E whole-slide pairs of 22 patients (Mendeley says CC BY 4.0): aligned by construction, so a known warp, which fidopt does not draw on large images yet (or tiles, or people's points); waits for a known warp on large (OME-Zarr) images
same section, unstained vs stained brightfield: weak contrast in the unstained image makes it a hard multimodal casetwo people click 20 landmarks (hair follicles, gland ducts, epidermal folds) on 10 pairs
EMIT (Exemplar Microscopy Images of Tissues) CyCIF TMACyCIF on a TMA of 34 cancer, non-neoplastic and normal tissue cores, nuclear stain per cycleCores x cycles (cycle pairs per core)None for registration (curated segmentation/cell types)one imageaccountCC BY-NC-ND (per RRID record); availability marked restrictednon-commercialneeds an account
the download needs an account
Whether raw unregistered cycles are included is unverified; ND licence limits derived benchmarksNuclear centroid matching across cycles verified by a person
EPFL mouse liver/intestine serial sections15 H&E serial sections14 consecutive pairsNoneone imagedirectCC BY 4.0opentruth to make
15 serial H&E sections (14 pairs) of mouse liver/intestine, CC BY 4.0, 3 GB, no truth; one or two specimens
Truth failsTwo people place 15 landmarks (vessels, crypts, gland points) on all 14 consecutive pairs.
fDNC C. elegans neuron correspondence datasets (Yu 2021)Whole-brain calcium imaging of moving worms (NeuroPAL and GCaMP); neuron point clouds across time and across animalsMany frame-to-frame and worm-to-worm pairs (counts not checked)Human-annotated neuron identities and tracks on both point setsboth imagesdirectnone statedaskwaits for a tool
neuron point sets with identities, no images: unchanged; waits for mesh/point-cloud scoring
Point-set registration with exact identities; images and point clouds Licence checked on 2026-10-10 at https://osf.io/t7dzu/; derived test pairs shareable: unclear; commercial use: unclear. OSF project is public but node_license is empty; the eLife article itself is CC BY 4.0
Fiducials for SXT and SIM correlation series (EMPIAR-10617 to 10622)soft X-ray tomography with six fiducial types (Dynabeads, gold, nanodiamonds, lysosomes, lipid droplets, nucleus) for SIM correlationone entry per fiducial typefiducials visible; matching SIM data unverifieddirectCC0 (EMPIAR default, unverified)openleft out
only soft X-ray tilt series are deposited and no matching SIM data is linked or found, so no second image
fails the pair criterion unless the SIM volumes are deposited elsewherefiducial centroids in SXT and SIM as landmarks if SIM is found
Flysta3D Drosophila 3D Stereo-seqStereo-seq of serial cryosections (7-10 um) of whole Drosophila embryos (14-18 h) and larvae, combined into 3D transcriptomesConsecutive sections per sample (counts not stated)Authors' 3D reconstruction; tissue annotationsboth imagesdirectnone stated on STOmicsDB pagesasktruth to make
tissue labels per section, no points; ssDNA image release and licence unverified (ask)
Whether ssDNA images are downloadable is unverified Licence checked on 2026-10-10 at https://db.cngb.org/stomics/flysta3d/; derived test pairs shareable: unclear; commercial use: unclear. Flysta3D site names no licence; CNGBdb terms page not readable without JavaScriptTissue labels on both sections; add landmarks for a TRE
Fuse My Cells (ISBI 2025, single view vs fused multiview light-sheet)3D light-sheet microscopy: one view and the multiview fused image of the same sample (nuclei, membranes)400+ paired 3D imagespairs share one frame by construction (fusion is registered to the views)both imagesaccountCC BY (rules page)openneeds an account
the download needs an account
Use is aligned-by-construction pairs with anisotropic vs isotropic appearance.apply known smooth warps to the aligned pairs for exact truth; if raw individual views are shared, stage rotation angles give multiview registration truth
Glioblastoma MALDI-FTICR MSI cohort with H&E and histology-to-MSI transforms (Mohammed, Abu Sammour 2026)MALDI MSI (FTICR) of glioblastoma sections with H&E, unstained brightfield and pixel labelsone directory per section (about 30 files; section count unverified)histology-to-MSI transformation parameters (elastix with point metric, Mannheim group) and pathologist pixel labelsone imagedirectCC BY 4.0opentruth to make
29 MSI sections of 18 patients with H&E (.svs), pre-MSI bright-field and a transformation matrix (.xlsx) per section (Zenodo listing), CC BY 4.0, 17 GB; no point files, labels on H&E only; MSI images to render from imzML
H&E from adjacent sections per the 2023 methods; brightfield image taken before MSI is intrinsically registered to MSIbrightfield-to-MSI is exact by acquisition; H&E-to-brightfield needs 20 landmarks by two people
Gold fiducials in cryo-EM/ET images (Zenodo 15830961)cryo-electron images of vitreous samples with gold beads; tilt-series alignment, same sample at many tilt anglesimages from 8 EMPIAR entries and 1 CryoET Data Portal set; full tilt series in EMPIAR and the portalmanual gold bead masks (napari) on single imagesone imagedirectCC BY 4.0openwaits for a tool
projection images of tilt series: unchanged; waits for 2D/3D pose error
Images rescaled to 8 A per pixel. Tilt-series alignment is a projection registration task, unusual for the benchmark.detect beads on every tilt (masks train the detector), track them across the series; IMOD fiducial-model residuals (about 1 px) give the alignment truth; limit: beam-induced bead motion
HANCOCKHead and neck cancer TMA cores stained on separate (serial) slides with H&E + 7 IHC markers (CD3, CD8, CD56, CD68, CD163, PD-L1, MHC-1), plus H&E WSIs763 patients, 2 cores each x 8 stainsTumour region annotations on WSIs; no cross-stain landmarksone imagedirectCC BY 4.0opentruth to make
TMA cores of 763 patients in H&E and 7 IHC on separate slides, per-core PNGs (12,236 files, 197 GB) direct from hancock.research.fau.eu, CC BY 4.0: the best ACROBAT-like candidate once two people place points on 10+10 patients
Fails truth; ACROBAT-like multi-stain registration material.Two people place 10 landmarks (gland and stroma corners, vessels) on the H&E and each IHC core of 20 patients.
HE-VS-MPM (TCIA)H&E, CK5/6 IHC and multiphoton microscopy of breast DCIS12 patientsNoneone imagedirectCC BY 4.0opentruth to make
H&E, IHC and multiphoton of 12 patients, CC BY 4.0, 43 GB, no truth; if the MPM is of the H&E section itself, a known warp instead
Truth failsIf MPM was imaged on the H&E section itself, the two are aligned by construction and a known warp gives exact truth; otherwise two people place 10 landmarks on each pair of the 12 patients.
HEMIT (H&E to multiplex IHC, same section)Same tissue section stained first with mIHC (DAPI, CD3, panCK) then bleached and H&E stained; colon cancer, 8 WSIs from the ImmunoAIzer cohort5292 aligned 1024x1024 patch pairs (3717 train, 630 val, 945 test)Patches registered by the authors with a 2-step registration (cellular-level alignment claimed); no landmarksone imagedirect: Mendeley Data, one HEMIT.7z of 10.6 GB (not readable by ranges), no accountCC BY 4.0 (Mendeley Data)openleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): open, but not among the six minor sets kept. It was proposed as: 5,292 pairs of 1024 x 1024 TIF patches (3,717 train, 630 validation, 945 test), H&E and three-channel mIHC (DAPI, panCK, CD3) of the same section, registered at the cellular level by the authors, from 8 whole slides of colon cancer patients (ImmunoAIzer cohort). Pairs are aligned by construction, so the corpus would warp one side by a known deformation (synthetic), and 8 patients give fewer than ten independent sources; CC BY 4.0, one 10.6 GB .7z on Mendeley Data. No patient key is released: the patch names carry no slide or patient, so cases cannot be grouped by patient.
Already-aligned patches only (repo github.com/BianChang/HEMIT-DATASET); no raw misaligned WSI pair, so only usable with a synthetic warpPairs aligned by construction: apply a known smooth deformation to one image to get exact multimodal truth (DAPI vs hematoxylin)
HER2-positive breast cancer spatial transcriptomics serial sections (Andersson 2021)Spatial transcriptomics with H&E images, HER2+ breast tumours, 8 patients with 3 or 6 serial sections each~30 adjacent-section pairs (36 sections)Pathologist region annotations on one section per patient; spot-to-pixel coordinate files for all sectionsone imagedirectCC BY 4.0 (files password-protected, password given on the record)opentruth to make
pathologist regions on one section per patient only (8 patients, 36 sections); annotating the adjacent section the same way would give labels on both
Annotations exist on a single section per patient, so label truth is one-sidedAnnotate the adjacent section with the same pathologist label scheme, or two people place 15 landmarks per pair
HER2match (same-slide H&E then HER2 IHC restain)breast cancer sections stained H&E, washed, restained HER2 IHC; same section17 whole-slide pairs; 21,172 curated tile pairsregistration by DeeperHistReg non-rigid (2 slides with Warpy), by the authorsboth imagesdirectCC BY-NC-SA 4.0non-commercialleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): its licence is not shown to allow commercial use (CC BY-NC-SA 4.0). It was proposed as: 17 same-section H&E/HER2 slide pairs (0.25 um) plus 21,172 tiles registered by the authors (Zenodo), CC BY-NC-SA: a known warp on the tiles now, on the slides once warps work on large images
restained same section: true deformation is small and near-rigid, so whole slides give near-exact truth after a careful fitrestained section: fit a careful landmark-based transform on 20 nuclei per slide, then apply a known warp for exact truth
HEST-1k (spatial transcriptomics with aligned H&E WSIs)1,229+ spatial transcriptomics profiles (Visium, Visium HD, Xenium, ST) each linked and aligned to an H&E whole-slide image, many organs, human and mouse~1,200 ST-grid/H&E pairs; Xenium entries give DAPI-to-H&E pairsSpot or cell coordinates in WSI pixel space, aligned by the authors (fiducial detection, keypoints) or by 10xone imageaccountCC BY-NC-SA 4.0non-commercialneeds an account
the download needs an account
Gated on Hugging Face (accept terms); alignment quality varies by sourceUse provided alignments as reference; verify with 10 landmarks on a 30-sample test subset
Hierarchical etched grid for correlative SXRF and optical imaging (Warwick)synchrotron X-ray fluorescence maps vs optical images of tissue on a finder gridcount unverifiedetched grid marks visible in both modalitiesdirectCC BY 4.0openleft out
a handful of demonstration ROIs (80 MB archive); the etched grid fixes only a rigid frame
physical markers by designgrid crossings picked by a program and checked by a person
HistologyHSI-BC-Recurrence (TCIA 2025)H&E whole-slide images and hyperspectral bright-field microscope images of the same slides, breast cancer47 WSIs and 677 hyperspectral fields to place in themnone for alignment; pathologist annotations on WSIsone imagedirectunverified (TCIA collection licence)unknowntruth to make
677 hyperspectral fields to place in 47 H&E slides: a localisation problem, no alignment truth, TCIA licence unverified
Field-of-view localisation, RGB vs hyperspectral.render pseudo-RGB from the hyperspectral cube, a program matches it to the WSI and a person verifies; two people mark 10 cell landmarks on 30 fields
HNSCC-mIF-mIHC-comparisonHead and neck carcinoma: the same tissue section stained by multiplex immunofluorescence, then restained by multiplex IHC (same section, two stains)8 patients, 9 ROIs each (512x512 patches)Co-registered mIF/mIHC ROI pairs from the authors; cell annotationsboth imagesdirect: IBM Aspera public link on the TCIA page, no accountCC BY 4.0 (TCIA page)openleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): open, but not among the six minor sets kept. It was proposed as: Version 2 (2023-08-31, CC BY 4.0, 1.01 GB of PNG on an Aspera public link) holds the same tumour sections of 8 head and neck carcinoma patients stained by multiplex immunofluorescence then restained by multiplex IHC, nine ROIs a slide cut into 512 x 512 patches that the page says are 'now perfectly co-registered'. Aligned by construction, the pairs need a known warp from the corpus (synthetic truth), and 8 patients give under ten test pairs: ranked apart.
Ships already-registered patches; raw unregistered pairs not confirmed.
HTAN CRC 3D atlas CRC1 serial sections (Lin, Wang, Coy 2023 Cell)Serial sections of one colorectal cancer specimen, alternating H&E and 20-plus-plex CyCIF, for 3D reconstruction; plus 16 other CRC specimens (CRC2-17)CRC1: 47 imaged sections (H&E and CyCIF) of a ~106-section block; adjacent section pairs and a full 3D stackNone released for registration; authors' own 3D reconstruction and cell tablesone imagedirectCC BY 4.0 (HTAN open-access imaging data)opentruth to make
47 serial H&E/CyCIF sections of CRC1 plus CRC2-17, public S3 (anonymous listing works), CC BY 4.0, 2.1 + 4.4 TB; no truth; CyCIF cycles share DAPI
Related to but larger than the listed HTAN-TNP-SARDANA pair (same HTAN CRC programme); DAPI in every CyCIF cycle also gives within-section cycle pairs Licence checked on 2026-10-10 at https://humantumoratlas.org/data-access; derived test pairs shareable: yes; commercial use: yes. Zenodo index record 10.5281/zenodo.7554924 (licence 'other-at') says all images are released via HTAN; HTAN data access page: open-access imaging data in IDC and SB-CGC are CC BY 4.0Two people place 20 landmarks (vessels, gland centres) on 10 adjacent pairs; within each CyCIF section, cycle-to-cycle DAPI pairs give near-exact truth after the authors' ASHLAR alignment
HTAN imaging collections in IDC (HMS, OHSU, Vanderbilt, WUSTL)Multiplexed IF (CyCIF, MxIF and others) and H&E whole-slide images of the same specimens: melanoma/colon (HMS, 16 subjects), breast (OHSU, 1), colon precancer (Vanderbilt, 30), pancreatic cancer (WUSTL, 21)Per specimen H&E vs mIF sections (same or adjacent, varies by centre); counts per collection unverifiedNoneone imagedirectCC BY 4.0 (IDC Zenodo pages 22845785, 22837342, 22837340, 22845786)opentruth to make
H&E and multiplexed IF slides in IDC (CC BY 4.0), same or adjacent sections by centre, no truth
DICOM Slide Microscopy; section adjacency must be checked per specimen; HTAN-TNP-SARDANA already listedTwo people place 20 landmarks on 10 H&E-mIF pairs; DAPI vs hematoxylin nuclei where sections are the same
HTAN-TNP-SARDANA (IDC)Colorectal cancer specimen: one H&E section and one adjacent CyCIF section (~20 markers)1 specimen, 1 pairNoneone imagedirectCC BY 4.0openleft out
one H&E/CyCIF pair: a demo, not a test set
Fails truth; single pair. IDC-only, published 2026-10-07.One pair: two people place 20 landmarks on the H&E and the CyCIF DAPI; CyCIF channels of the same section under a known warp give exact multi-channel truth. A demo case, not a test set.
HuBMAP Florida lymphoid organs: CODEX, 2D and 3D IMC and light-sheet (lymph node, spleen, thymus)CODEX, imaging mass cytometry (2D and 3D serial sections) and light-sheet of the same donors' lymph node, spleen and thymus61 CODEX, 13 2D IMC, 3 3D IMC, 3 light-sheet datasetsnoneone imagedirectHuBMAP: permissive, such as CC BY 4.0openneeds an account
HuBMAP downloads go through a Globus sign-in (HuBMAP's documentation; not tested on a file); serial 3D IMC sections, cross-modality pairs likely from other blocks; no truth
3D IMC sets are serial-section stacks (section-to-section registration); cross-modality pairs are same organ, likely different blocks; url is one example dataset, the group has many more (search the portal by group)serial IMC: vessel and follicle centres tracked across sections by a program, person verifies; cross-modality only label overlap of follicles
HuBMAP Rochester lung: autofluorescence, PhenoCycler and H&E on the same section (LungMAP)pre-stain autofluorescence, 27-45 plex PhenoCycler IF and H&E after IF on the same FFPE lung section, paediatric and adult donors26 autofluorescence, 8 PhenoCycler, about 205 histology datasets (not all same-section triplets)noneone imagedirectHuBMAP: permissive, such as CC BY 4.0openneeds an account
Globus sign-in to download; same-section autofluorescence, PhenoCycler and H&E of lung, no truth (nuclei matching would make one)
descriptions: 'H&E performed after multiplexed immunofluorescence assay on same section' and 'See H&E and phenocycler stained image data files of same section'; url is one example dataset, the group has many more (search the portal by group)nuclei matching DAPI vs haematoxylin; autofluorescence vs H&E alveolar wall landmarks clicked by two people
HuBMAP Stanford bone marrow cohort: MIBI, MALDI N-glycan IMS and H&EMIBI-TOF, MALDI N-glycan imaging and H&E of bone marrow cores from the same donors (same or serial sections unverified)46 MIBI, 39 MALDI, 40 H&E datasetsnoneone imagedirectHuBMAP: permissive, such as CC BY 4.0openneeds an account
Globus sign-in to download; MIBI, MALDI and H&E of bone marrow whose shared section is not stated, no truth
whether MIBI, MALDI and H&E share a section is not stated in the snippets seen; url is one example dataset, the group has many more (search the portal by group)bone trabeculae and fat cells as landmarks on H&E and MIBI/MALDI, two people, 15 points per pair
HuBMAP Stanford intestine CODEX with H&E after CODEX on the same section (Hickey 2023)CODEX 50+ marker immunofluorescence and H&E of the same small and large intestine sections, 8+ donorsabout 67 CODEX and 56 H&E datasets; descriptions say H&E performed after CODEX or H&E corresponding to CODEXnone; same section, alignment by the userone imagedirectHuBMAP: permissive, such as CC BY 4.0openneeds an account
Globus sign-in to download; CODEX then H&E on the same intestine sections (8+ donors), no truth: a strong same-section candidate once nuclei pairs are checked by a person
some H&E descriptions say 'corresponding' rather than 'after CODEX': check same vs adjacent section per dataset. Stanford also lists 20 Xenium small-intestine sections from a multi-section Snyder/Sorger collaboration (block B015-H); url is one example dataset, the group has many more (search the portal by group)same section: CODEX DAPI vs H&E haematoxylin gives near-exact nuclei correspondence; detect nuclei on both, match, a person checks 20 per pair
HuBMAP UCSD placenta: H&E, Xenium and Cell DIVE on the same sectionsH&E, Cell DIVE multiplex IF and Xenium spatial transcriptomics of 21 FFPE placenta sections across gestation, same section21 sections, three modalities eachnone releasedone imagedirectHuBMAP: permissive, such as CC BY 4.0openneeds an account
Globus sign-in to download; 21 placenta sections with H&E, Xenium and Cell DIVE of the same section (entity API: public, published), no truth
page states H&E, Xenium and Cell DIVE were generated from the same sections; good trimodal set; url is one example dataset, the group has many more (search the portal by group)nuclei from Xenium DAPI, Cell DIVE DAPI and H&E haematoxylin matched automatically, person verifies 20 per pair; same section so near-exact
HuBMAP Vanderbilt BIOMIC kidney: MALDI IMS, pre/post autofluorescence and PAS on the same sectionMALDI imaging mass spectrometry, autofluorescence microscopy and PAS histology of one kidney section (plus PhenoCycler/CODEX on some), same sectionabout 111 MALDI, 330 autofluorescence, 236 histology datasets from about 30 donors (kidney, also eye, pancreas)none released as truth; group's own wsireg/image2image alignment, IMS pixels aligned to laser ablation marks on post-IMS autofluorescenceone imagedirectHuBMAP external data sharing policy: permissive, such as CC BY 4.0 for open data; credit HuBMAPopenneeds an account
Globus sign-in to download; MALDI, autofluorescence and PAS of the same kidney sections, ablation marks give IMS points, no released truth
same section so the true transform is near rigid/affine plus small local warp; laser ablation marks are a physical grid visible in post-IMS autofluorescence. Download via Globus/HTTPS, login not checked; url is one example dataset, the group has many more (search the portal by group)ablation-mark grid detected on post-IMS autofluorescence gives exact IMS-to-microscopy points; two people click 20 glomeruli or vessel centres on autofluorescence and PAS for a test set
HuBMAP Vanderbilt eye: autofluorescence, H&E and PhenoCyclerautofluorescence microscopy (3 channels) and H&E brightfield of human eye sections, some with PhenoCycler multiplex IF, same donors (BIOMIC workflow images AF before staining)108 autofluorescence and 108 H&E datasets; 7 PhenoCyclernoneone imagedirectHuBMAP: permissive, such as CC BY 4.0openneeds an account
Globus sign-in to download; autofluorescence and H&E of eye sections, same section not stated, no truth
equal counts suggest AF and H&E of the same sections, but the descriptions do not say so (unverified); url is one example dataset, the group has many more (search the portal by group)if same section: retinal layer boundaries and vessel centres on AF and H&E by two people; or nuclei matching DAPI-channel AF vs haematoxylin
Human middle temporal gyrus Visium, Alzheimer's vs control (Chen 2022)10x Visium with H&E, human MTG, control and Alzheimer's disease donors; between subjectsControl and AD sections (counts not checked); used as heterogeneous integration pairsManual cortical layer (L1-L6) and white-matter labels per spot by the original authorsboth imagesdirectnone stated; data in GEO GSE220442 and scREADaskto buildpriority 3
manual layer labels on every Visium spot of both sections, but the sections come from different donors cut at different places, few donors, and GEO states no licence (ask)
Accession not confirmed (NCBI pages blocked here); layer labels like DLPFC, so same evaluation recipe Licence checked on 2026-10-10 at https://pmc.ncbi.nlm.nih.gov/articles/PMC9773466/; derived test pairs shareable: unclear; commercial use: unclear. Acta Neuropathol Commun 2022 data statement gives GEO GSE220442 and bmbls.bmi.osumc.edu/scread; GEO pages were captcha-blocked so GEO's general terms were not readLayer label transfer after alignment (like DLPFC); two people place 10 landmarks on layer boundaries for a point truth
HyReCohistology, 9 cases × 4 consecutive stains, plus 54 restained H&E/PHH3 pairs86 pairs690 + 2,303 landmarksboth imagesaccountCC BY-SA 4.0openneeds an account
the download needs an account
status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (WSI registration review 2025). Used in 3 of the 2D and microscopy registration papers read (Lotz 2021; DeeperHistReg 2024; CORE 2025).
IBDColEpi (H&E and CD3 colon biopsies, DataverseNO)FFPE colon biopsies of inflammatory bowel disease and controls: 140 H&E and 111 CD3 IHC whole-slide imagesPossible H&E-CD3 pairs from the same biopsies; pairing not documented (IDs anonymised)Epithelium annotations on every slideboth imagesdirectCC0 1.0openleft out
anonymised IDs do not link an H&E slide to the CD3 slide of the same biopsy, so no known pairs
May fail: anonymised IDs keep only stain and activity, so matching H&E to CD3 of the same biopsy may be impossible (unverified)If pairs can be recovered (ask authors), epithelium labels on both give Dice; add 20 landmarks on 10 pairs
IBEX 2D and 3D microscopy data (Radtke 2020-2022)Iterative Bleaching Extends multiplexity: same section imaged over several cycles with Hoechst/JOJO-1 nuclear marker in every panel; human spleen, liver, jejunum, mouse spleenSeveral experiments of 3-24 cycles each (e.g. 4-cycle human spleen); 2D and 3D z-stacksNone; authors' affine registration on the common nuclear channel (XTRegisterSameChannel / SimpleITK)one imagedirectCC BY 4.0 (Zenodo records, unverified per record)opentruth to make
3D confocal z-stacks of one section over 4 staining cycles (spleen and liver), Hoechst in every cycle, no truth: nuclei matched and checked by a person
Imaris .ims format; further records 4304786, 5244551 and the IBEX knowledge-base on ZenodoCommon nuclear channel across cycles: matched nuclei verified by a person; 3D stacks allow volume truth
IMC, serial two-photon tomography and H&E of mouse embryos (S-BIAD2843)imaging mass cytometry vs STPT block-face volume vs H&E of the same embryoscount unverifiednone stateddirectCC BY 4.0openleft out
two embryos (IMC and H&E, STPT of one): too few independent pairs, no truth
2D to 3D section placement problemtwo people mark 15 anatomical landmarks on sections and the matching STPT plane for 10 sections
Immunofluorescence and imaging mass cytometry on the same slide (Zenodo 7576005)whole-slide high-resolution IF vs IMC of the same tissue sectioncount unverifiednone released as truth; authors' own IF-IMC registration in the pipelinedirectCC BY 4.0openleft out
same Zenodo record (7576005) as the 'Dual-modality IF and IMC whole-slide data with adjacent H&E (Kim 2023 Cell Rep Methods)' row
same section: DNA channel in both, so nuclei centroids matched by a program and checked by a person give dense truth
Immunostained serial sections registration data (Lobachev et al., MedIA 2017)serial histological sections with immunostaining, section to sectionfull data of the paper (count unverified)evaluation material of the paper (type unverified)directCC BY 4.0opentruth to make
full data and evaluation in a 5.4 GB zip whose content and truth the record does not describe, CC BY 4.0; check the zip, else two people place points
same material also on a Bayreuth servertwo people mark 10 landmarks (vessels, cell clusters) per adjacent pair on 10 pairs
Intravital two-photon movement artefact videos (Soulet 2013)in vivo two-photon time-lapse of mouse brain, spinal cord, sciatic nerve and myenteric plexus with breathing/heartbeat motionsupplementary movies (88 MB total)noneone imagedirectCC BY (PLOS figshare default, unverified)openwaits for a tool
intravital movies, no truth: unchanged; waits for point tracking
Small sample of real physiological motion; truth must be madetwo people track 10 cell bodies or vessel branch points over 50 frames
ISBI 2012 EM segmentation stack (Drosophila VNC ssTEM)30 serial section TEM images (aligned), section-to-section30 training sections (+30 test)membrane labels per sectionboth imagesdirectnone statedaskin the corpuspriority 3
In the corpus: 10 test sections of the provider's test stack and 10 development sections of its training stack, each paired with a known in-plane warp of itself (as SNEMI3D); ranked apart in serial-em-synthetic. No licence stated (ask).
Classic small stack; pre-aligned Licence checked on 2026-10-10 at https://imagej.net/events/isbi-2012-segmentation-challenge; derived test pairs shareable: unclear; commercial use: unclear. challenge page offers the training and test data for download with no terms; images from Cardona et al. 2010 (PLoS Biol)known warps per section
Ivy Glioblastoma Atlas Project ISH and H&E sectionsGlioblastoma blocks: in situ hybridisation sections and adjacent H&E sections annotated for anatomic structures; 8-42 tumours depending on surveyMany ISH-H&E adjacent pairs per block (hundreds of genes x tumours)Anatomic structure annotations on H&E; Allen registration of ISH to H&E (unverified)one imagedirectAllen Institute Terms of Use (non-commercial)non-commercialtruth to make
structure labels on the H&E side only (adjacent ISH unlabelled), image download path unverified, Allen terms non-commercial
Clinical database part needs registration; image download path unverified Licence checked on 2026-10-10 at https://alleninstitute.org/terms-of-use/; derived test pairs shareable: yes; commercial use: no. terms list glioblastoma.alleninstitute.org; use, copy, distribute and derive for research or other noncommercial purposes; no commercial redistribution or commercial service without written permissionTransfer H&E structure labels to ISH and ask a person to correct them on 20 pairs; or 20 landmarks on 10 pairs
JRC2018 Drosophilafly brains, inter-specimenmanyevaluated with automatic neuron skeletonsboth imagesdirectCC BY 4.0 (Janelia figshare records)opentruth to make
3D fly brain templates and transforms on figshare, but not the paper's evaluation brains or skeletons: two people place 20 landmarks on 10 brain pairs
status change: the template paper is itself a registration benchmark with a released evaluation set and code (github.com/saalfeldlab/template-building) Licence checked on 2026-10-10 at https://www.janelia.org/open-science/jrc-2018-brain-templates; derived test pairs shareable: yes; commercial use: yes. transform records checked on figshare (11871699, 14991978, 24002001) are CC BY 4.0; template files sit behind figshare share links and a Janelia acknowledgement form, licence of those not readTwo people place 20 landmarks (neuropil corners, tract crossings) on 10 brain pairs; neuron skeleton overlap as a weaker check.
KPMP kidney biopsy imaging (Kidney Precision Medicine Project)Human kidney biopsies: whole-slide light-microscopy images (KPMP protocol stains, likely H&E, PAS, trichrome, silver on serial sections), plus open 3D cytometry, CODEX, IMC and multimodal imaging mass spectrometry with microscopyPer-biopsy stain-to-stain serial pairs; IMS-to-microscopy pairs (counts unverified)None for registrationone imagedirectno data licence named; public tier without DUA, controlled tier needs the KPMP DUAaskneeds a request
no data licence and reproduction needs KPMP's permission; EM and IF need the DUA
Stain list and seriality from protocol knowledge, unverified on page; EM and main-study IF need a DUA Licence checked on 2026-10-10 at https://atlas.kpmp.org/repository/; derived test pairs shareable: unclear; commercial use: unclear. repository is a JavaScript app (not readable here); earlier read of the site said reproduction needs permission; KPMP access rules (via search) put controlled data under a DUA; ask KPMPGlomeruli outlines on both stains (public kidney segmenters, then verified by a person) give labels; two people place 20 landmarks on 10 pairs
Langlieb 2023 Slide-seq whole mouse brain with adjacent Nissl sections and CCF alignment101 Slide-seq pucks of one hemisphere with Nissl-stained pre and post sections; Slide-seq to Nissl to Allen CCF101 pucks, each with 2 adjacent Nissl sectionsmanual fiducials in 3D Slicer for Slide-seq to Nissl (TPS), LDDMM to CCF; error 22-94 um by marker genes (fiducials released: unverified)both imagesdirectNeMO open access (terms unverified)unknownneeds a request
NeMO gives no access, licence or links for the 6 TB analysis record, the manual fiducials are not in any listing, and a puck is a bead point cloud rather than an image
also at BrainCellData.org and Broad Single Cell Portal SCP2157if fiducials are not released: two people click 15 landmarks puck vs Nissl on 10 pucks; region-marker genes give a label check
Learn2Reg COMULIS3D CLEM and SHG to bright field40 test~25 manual landmarks a case, validation and test withheldboth imageshiddenCC BY 4.0 (SHG/BF images)openleft outThe 3D-CLEM task was dropped. status change: Round 4 count: listed in 2 of the 38 round-4 surveys and benchmark papers (Chen 2025 survey (MedIA); Learn2Reg 2024 overview). Used in 2 of the 2D and microscopy registration papers read (Learn2Reg 2024; FIDE 2024).
LICONN expansion light-microscopy connectomics (Tavakoli 2025)Hydrogel-expanded mouse brain tissue imaged by spinning-disk/light microscopy at synapse resolution, tiled volumes with molecular channels; same tissueTile-to-tile; channel-to-channelNone for alignment; dense neurite reconstructions by the authorsone imagedirectCC BY-NC-SA 4.0non-commercialtruth to make
3D expansion-microscopy volumes (TIFF up to 3 GB, larger ones in zarr/n5), no alignment truth: a known warp between two channels of one volume
Weak for registration; useful as large ExM test volume.Tile overlaps and stage positions for stitching truth; known warp per chunk; neurite traces as a downstream check
Light and electron microscopy of 3D cell cultures (EMPIAR-11380)light microscopy vs FIB-SEM volumes of human organoidscount unverifiedFIB-SEM segmentations; no LM-EM correspondences stateddirectCC0 (EMPIAR default, unverified)opentruth to make
3D confocal stacks of 2 organoids (S-BIAD610) and their FIB-SEM volumes (EMPIAR-11380), no correspondences: two people mark nuclei or lumen points
two people mark 10 nuclei or lumen landmarks in LM and EM
Light My Cells database (BioImage Archive S-BIAD1047)transmitted light (bright-field, phase, DIC) and fluorescence images of the same cells, many labs2,574 acquisition sets, 56,984 2D imagespairs aligned by acquisition (same field, same instrument)both imagesBioImage Archive, direct per file (about 5 to 7 MB an image)CC BY 4.0 (BioImage Archive study)openin the corpuspriority 3
56,981 OME-TIFF images from 30 studies of 8 French imaging centres: transmitted-light images (bright-field, phase contrast or DIC, z-stacks) each paired with fluorescence of the nucleus, mitochondria, tubulin or actin, acquired on the same field (BioImage Archive study). In the corpus as a synthesis dataset (LightMyCells_Nucleus_Syn); a registration set of these pairs under a known warp would derive from it.
Sci Data 2026. Multimodal pairs aligned by construction.apply known smooth warps for exact multimodal truth
Lobachev serial sectionsmicro-CT, SBF-SEM, light-sheet lung stacks with simulated sectioning distortion600 / 1,000 / 300 slicesknown distortion maps (synthetic)both imagesdirect: Zenodo record 4282448; 7z archives, not readable by rangesCC BY 4.0 (Zenodo)openleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): open, but not among the six minor sets kept. It was proposed as: Zenodo (CC BY 4.0, 14.2 GB of 7z archives): three lung stacks cut into 2D sections, a rabbit lung in micro-CT (600 to 700 slices), a mouse lung in SBF-SEM (1,000) and a rat lung in light-sheet (300 to 400), each as undistorted ground truth and with the provider's synthetic local distortions and rigid moves (IEEE Access 2021). The distorter is public (GitHub), so the corpus would draw its own private warp on the ground-truth sections (synthetic); three specimens only, ranked apart.
Macaque cortex Stereo-seq single-cell atlas (Chen 2023 Cell)Stereo-seq of 161 coronal sections from 3 cynomolgus monkeys, with adjacent immunostained sections; section-to-section and to atlas119 sections (monkey 1, 500 um spacing) plus 19 and 23none released as truth; cortical region labelsone imagedirectunverifiedunknowntruth to make
Stereo-seq chip images of 161 sections of 3 monkeys, very large, no truth, licence unverified
centimetre-scale sections; nucleic acid and ConA staining on the chip give an image for each sectioncortical layer boundaries from marker genes on adjacent sections as label truth; sulci landmarks by two people
MaLissa (Lissajous confocal endomicroscopy)Porcine stomach ex vivo, 10 Hz sparse Lissajous CLE frames matched to a mosaic of slow high-quality frames; frame-to-mosaic and frame-to-frame alignment2,060 LQ-HQ pairs from 16 clips (3,988 frames)HQ crop chosen per LQ frame by FFT phase correlation (translation only), algorithm-madeone imagedirectCC BY-NC 4.0 (HF API)non-commercialleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): its licence is not shown to allow commercial use (CC BY-NC 4.0 (HF API)). It was proposed as: frame-to-mosaic crops chosen by phase correlation: translation only, nothing deformable after alignment; non-commercial
Weak, translation-only truth; animal tissue.Apply known sub-pixel shifts and small warps to HQ mosaic crops and resample with the Lissajous mask: exact truth.
mapzebrain larval zebrafish brain atlas (Kunst 2019)Confocal stacks of transgenic and antibody-stained larval zebrafish brains registered to a common reference; between fishMany lines registered to one reference (counts unverified)Region annotations on the reference (unverified)one imagedirectnone statedasktruth to make
3D confocal larval brains in one reference space; the JavaScript site could not be read for raw stacks or labels: two people place 15 landmarks on 10 fish
Page content not readable by the tools; details unverified. Z-Brain and ZAPBench already listed Licence checked on 2026-10-10 at https://mapzebrain.org/; derived test pairs shareable: unclear; commercial use: unclear. site is a JavaScript app; its bundle contains no licence or terms textTwo people place 15 landmarks (commissures, ventricle corners) on 10 fish and the reference
mBrainAligner sample whole-brain data (fMOST, LSFM, VISoR, MRI)Whole mouse brains in 4 modalities to Allen CCFv3; cross-modal, between subjects1 downsampled (~50 um) brain per modality plus a partial LSFM brain; full-res fMOST brains on BILNone released with the samples; no landmark or label files mentioned on the repo pageone imagedirectslightly revised MIT (repo licence covering mBrainAligner and associated files)opentruth to make
4 downsampled 3D whole mouse brains (fMOST, LSFM, VISoR, MRI) and the CCF, only network region maps and registration priors: two people place the standard CCF landmarks
Full-resolution fMOST at download.brainimagelibrary.org/biccn/zeng/luo/fMOST/ Licence checked on 2026-10-10 at https://github.com/Vaa3D/Vaa3D_Wiki/wiki/License_mBrainAligner; derived test pairs shareable: yes; commercial use: yes. the ~50 um sample brains sit in the repo's examples folder under the repo's MIT-style licence (adds a no-endorsement clause); a software licence applied to data; full-resolution LSFM and VISoR only on requestTwo people place the 18-27 standard CCF landmarks (as in D-LMBmap and Gubra) on each sample brain
MCMICRO exemplar-001/002 raw CyCIF cyclesCyclic immunofluorescence (CyCIF): same tissue imaged over 10 cycles, Hoechst/DAPI in every cycle; exemplar-001 a tissue region, exemplar-002 a 4-core TMA10 unregistered, unstitched cycles per exemplar (45 cycle pairs each), multi-tile OME-TIFFNone; ASHLAR stitching and registration output is the de facto referenceone imagedirectnone stated for the S3 data; mcmicro.org site content CC BY-NC 4.0; code MITaskleft outpriority 3
two exemplars (one specimen and a 4-core TMA) of raw unstitched tiles: a demo, not a test set
Small, fast, raw tiles: ideal test of tile stitching plus cycle-to-cycle rigid/affine registration Licence checked on 2026-10-10 at https://mcmicro.org/datasets/; derived test pairs shareable: unclear; commercial use: unclear. public S3 bucket carries no licence; the datasets page footer licenses 'contents of this site' CC BY-NC 4.0, which may or may not cover the images; ask LSPDAPI repeats every cycle: nuclei centroids matched across cycles and verified by a person give dense, near-exact truth; or ask two people to place 20 nuclear landmarks per cycle pair
MERFISH mouse hypothalamic preoptic region (Moffitt 2018)MERFISH single-cell spatial transcriptomics, mouse hypothalamus, serial coronal slices at known Bregma positions, several animalsConsecutive slices per animal (benchmarks use 5 slices, 4 pairs)Cell-type and structure annotations per slice; slice z-positions knownboth imagesdirectCC0 1.0openwaits for a tool
cell positions without images: unchanged; waits for mesh/point-cloud scoring
Dryad holds cell coordinates and expression; raw images not included (unverified) Used in 2 of the 2D and microscopy registration papers read (SLAT 2023; RAFT-UP 2026).Structure labels on both; known slice spacing gives a z-check
MERFISH mouse primary motor cortex 33-slice series (via SPACEL data)MERFISH, mouse primary motor cortex (Zhang et al. 2021), 33 consecutive slices, for 3D reconstruction32 consecutive slice pairsCortical layer annotations (L1-L6, WM) per cellboth imagesdirectCC BY 4.0openwaits for a tool
cell point clouds with layer labels, no images: unchanged; waits for mesh/point-cloud scoring
Cell point clouds rather than images; original images on the Brain Image Library (unverified)Layer labels on both; add cell-type landmark clusters
METASPACE public MSI datasets with optical imagesMALDI/DESI/other imaging mass spectrometry of tissues, many with an optical (brightfield or H&E) image of the same section uploaded by the submitterthousands of public datasets; share with optical images unverifiedoptical image placed on the ion images by the submitter (manual alignment in the web tool, unverified)one imagedirectnot stated on the terms page (unverified)unknowntruth to make
submitters' optical overlays of uneven quality, licence not stated; 20 datasets with post-MSI H&E would need people's points
huge pool for MSI-to-microscopy pairs; alignment quality varies by submitterpick 20 datasets with post-MSI H&E; ablation marks or tissue edges as landmarks by two people
MIIT serial-section multi-omics test data9 serial histology sections (HES, HE, MTS, IHC) with Visium spots and MALDI-MSI on some sections9 sections, adjacent and non-adjacent pairslandmarks with matching labels across all sections, plus semi-manual tissue masks (authors)both imagesdirectCC0openleft out
same Zenodo record (14931377) as the 'MIIT test data: serial sections with landmarks, Visium and MALDI (Wess 2024)' row
reduced test data; MSI reduced to a few ions; who placed the landmarks is unverifieda second person re-places 10 landmarks per pair to measure observer error
MIIT test data: serial sections with landmarks, Visium and MALDI (Wess 2024)9 serial prostate sections (HES, HE, MTS, IHC), one with Visium, two with MALDI-MSI; between serial sections9 sections, 8 consecutive pairs plus any pairlandmarks with matching labels across all sections, plus tissue masks, by the authorsboth imagesdirect: Zenodo record 14931377, one tar.gz (not readable by ranges, small)CC0 1.0 (Zenodo)openleft outpriority 2
All sections belong to one prostate specimen. Retained as an example outside the benchmark in datasets/examples/ProstateSerialSections.
reduced test data (MSI only a few m/z); rare public multi-stain serial set with landmarks on every section
MIMIC IMC-MSIImaging mass cytometry vs MALDI-MSI, same slideLiver biopsies, TMAsNoneone imagedirectCC0opentruth to make
same-slide IMC and MALDI (liver biopsies, TMAs, CC0), no truth; the resolution gap limits people's points
Truth failsSame slide: two people place 10 landmarks (tissue edges, core features) on IMC and MSI of 10 regions; the resolution gap limits accuracy.
MIPHEI-ViT processed ORION-CRC and HEMIT tilesAligned restained H&E and mIF tiles derived from ORION-CRC and HEMITTile pairs from 40 ORION specimens and HEMIT patchesAlignment from the source datasets plus nucleus segmentationone imagedirectMITopenleft out
tiles derived from ORION-CRC and HEMIT, whose own rows hold the images (no image twice)
Derivative of ORION-CRC and HEMIT; convenient tile formatAligned by construction: apply known deformations to tiles for exact truth
MIST (Multi-IHC Stain Translation)Adjacent sections, H&E vs HER2, Ki67, ER, PR IHC, breast cancer~5000 patch pairs per stain (1024x1024, 20x), from 56-64 WSIsPatches pre-aligned by the authors; misalignment acknowledgedone imagedirectnot stated for the data; code BSD-styleasktruth to make
adjacent-section patches pre-aligned with acknowledged misalignment, whole slides not released, no data licence (ask)
Source WSIs not released; four stains match ACROBAT's (origin unverified) Licence checked on 2026-10-10 at https://github.com/lifangda01/AdaptiveSupervisedPatchNCE; derived test pairs shareable: unclear; commercial use: unclear. README gives Google Drive/Baidu links with no data terms; repo LICENSE is a BSD-style code licenceTwo people place landmarks on a 50-pair subset per stain
MIST sample tile grids (NIST)5x5 and 10x10 tile grids (phase contrast and Cy5), 10% overlap2 grids x 2 modalitiesnone beyond nominal stage grid; example MIST resultsone imagedirectNIST work, not subject to US copyright (17 U.S.C. 105)openleft out
small demo grids with nominal stage positions only; the NIST reference stitching data (another row) is the same kind of set with measured positions
Small demo sets; truth only in the larger reference data Licence checked on 2026-10-10 at https://www.nist.gov/open/copyright-fair-use-and-licensing-statements-srd-data-software-and-technical-series-publications; derived test pairs shareable: yes; commercial use: yes. sample tiles ship in the usnistgov/MIST wiki; NIST policy: data by NIST employees is generally not copyrighted in the US (may be abroad), acknowledge NISTuse the reference colony data, or crop overlapping tiles from one large image (exact offsets)
MITOS-ATYPIA-14 (same slides on two scanners)H&E breast cancer biopsies scanned on Aperio ScanScope XT and Hamamatsu NanoZoomer 2.0-HT284 x20 frames and 1136 x40 frames per scanner; same slides (frame-level correspondence unverified on page)Mitosis and atypia annotations per frame (pathologists), not registration truthboth imagesaccountnone statedaskneeds an account
the download needs an account
Cited by Wilm 2023 as having local image correspondences across scanners; small near-rigid pairs Licence checked on 2026-10-10 at https://mitos-atypia-14.grand-challenge.org/Donwload/; derived test pairs shareable: unclear; commercial use: unclear. Dataset and Download pages give pCloud links and a data description PDF, no licence or terms; slides from Pitie-Salpetriere (Paris); ask the organisersMitosis point annotations on both scanners' frames act as landmarks; add 10 nuclei landmarks per frame pair
MOSTA mouse embryo Stereo-seq (Chen 2022 Cell)Stereo-seq spatial transcriptomics with ssDNA images, mouse embryos E9.5-E16.5, sections across stages and some consecutive sectionsConsecutive-section and cross-stage pairs (e.g. E11.5 vs E12.5 used in benchmarks); counts unverifiedRegion/organ annotations on each section by the authorsboth imagesdirectnone stated on STOmicsDB pagesaskleft out
same data as the 'MOSTA Stereo-seq mouse embryo atlas' row
Cross-stage pairs are between-subject and developmental; consecutive-section pairs are the registration-relevant ones Used in 3 of the 2D and microscopy registration papers read (SLAT 2023; Spotscape 2025; RAFT-UP 2026 (midbrain E12.5-16.5)). Licence checked on 2026-10-10 at https://db.cngb.org/stomics/datasets/STDS0000058; derived test pairs shareable: unclear; commercial use: unclear. dataset record (raw data CNP0001543) and resource page name no licence; CNGBdb terms page not readable without JavaScriptOrgan labels on both; add landmarks on ssDNA images for consecutive pairs
Mouse kidney MSI and H&E pair with pen-marked landmarks (Foll, Galaxy MSI)tryptic peptide MALDI image (150 um) and H&E of a mouse kidney FFPE section1 pairxylene-resistant pen crosses outside the tissue as landmarks, visible in bothboth imagesdirectCC BY 4.0openleft out
one MSI/H&E pair (maldi.png, he.png), no coordinates for the pen crosses: a smoke test, not a test set
toy-size single pair, useful as a smoke test
Mouse V1 two-photon imaging with Xenium of the same neurons (DANDI 002015)in vivo two-photon mesoscope planes and z-stack vs ex vivo Xenium serial sections, same neuronsmice with 4 sessions each (count unverified)authors registered Xenium cells to the in vivo z-stack and ROIs; cross-session cell IDsboth imagesdirect: DANDI draft version, NWB-zarr assets on the public S3 bucket, arrays readable one by one; no accountCC BY 4.0 (DANDI metadata: spdx:CC-BY-4.0)openleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): open, but not among the six minor sets kept. It was proposed as: The 36 NWB files (9 mice, 4 sessions each, 128 GB) hold per plane only the average and maximum projections, the ROI masks of an automatic segmentation and a cross-session cell id; the Xenium images and the cortical z-stack named in the row are not in the deposit, only Xenium cell tables matched to ROIs. What can be registered is one imaging plane across two sessions of a mouse, 2D and close to rigid, scored on ROI masks matched by the authors' pipeline; 9 mice cannot give ten test and ten development subjects. CC BY 4.0, open DANDI draft (no published version yet).
cell matches come from the authors' registrationtwo people verify 50 cell matches per mouse
MRI-histology coregistration with blockface (Dryad 2026)Specimen MRI vs histology sections via blockface, several species/regionsSeveral specimensAuthors' registrationdirectCC0opentruth to make
one human cortex block: specimen MRI and a 3D blockface volume (plus 2D histology slides) with only the authors' registration, no landmarks; two people would place landmarks on blockface and MRI
Two people place 10 landmarks on 10 section/blockface pairs and on blockface vs MRI; the authors' registration is a reference only.
Multi-modal WSI pipeline example: multi-cycle fluorescence, brightfield and H&E (SW480 / GR1000426)cyclic fluorescence with paired brightfield tiles per cycle and a final H&E of the same sample; tile stitching, cycle alignment, multimodal registration2 cycles plus H&E for one sample (example data)authors' reconstructed and registered images includedone imagedirectCC BY 4.0openleft out
one sample (two cycles and an H&E): a pipeline example, not a test set
example data of a pipeline paper; covers three registration sub-tasks in one samplenuclei from fluorescence vs brightfield phase retrieval vs H&E, person verifies
Multi-scanner canine cutaneous SCC dataset (Wilm 2023)Same H&E slides scanned on 5 scanners (Aperio CS2, NanoZoomer S210, NanoZoomer 2.0-HT, Pannoramic 1000, Aperio GT450), canine squamous cell carcinoma44 slides x 5 scanners (10 scanner pairs per slide, 440 pairs)Polygon annotations transferred between scanners by the authors' WSI registration (Marzahl quad-tree) and visually validatedboth imagesZenodo, direct, one TIFF per scanCC BY 4.0 (Zenodo)openleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): open, but not among the six minor sets kept. It was proposed as: 44 canine SCC slides (the CATCH SCC subset) each scanned on five scanners, released as 4 µm/px pyramidal TIFFs (220 images, 12.1 GB); the 1,243 CS2 polygons (tumour and six tissue classes) were carried to the other four scanners by the authors' quad-tree WSI registration and checked visually by overlay (Zenodo, arXiv 2301.04423). The same glass slide on two scanners is rigid after alignment, apart from stitching that is under a pixel at 4 µm, so the polygons would score agreement with the authors' registration; a known warp on the aligned scans (synthetic) is the truth that leaves room.
Released at reduced resolution (4 um/px) pyramids; near-rigid problem with scanner colour shiftSame physical slide: two people place 20 landmarks per pair on 10 slides; or keypoint matching verified by a person
Multimodal optical microscopy of cytoskeleton (HeLa, 5 modalities)bright-field, RICM, widefield fluorescence, TIRF and confocal images of the same HeLa cells2,253 cells, 11 channelsunverified whether modalities are pre-aligneddirectunverifiedunknownleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): its licence is not shown to allow commercial use (unverified). It was proposed as: five modalities of the same HeLa cells: a known warp if they share the field (record not opened, licence unverified)
Sci Data 2026. Data record not opened.if channels share a stage position, stage coordinates give alignment; otherwise two people mark 10 cell landmarks on 30 cells
MultiModal Platy-Browser (Platynereis, BioImage Archive S-BIAD14)whole-body serial block-face EM of a Platynereis larva vs gene expression atlas, between specimens1 EM volume vs atlaspublished registration of the atlas to EM (landmark-based, unverified)directunverifiedunknowntruth to make
3D: the record holds the ProSPr atlas reference and the EM volume at 0.55 um (one larva against the atlas) but no landmarks or transform, and no licence; nuclear landmarks by two people would make it
no licence on the recordtwo people mark 20 nuclear landmarks in EM and atlas
MUSCLEMOTION artificial-cell and cardiomyocyte moviesvideos of beating cardiomyocytes, cardiac organoids, zebrafish hearts; two artificial cells with programmed contractionsupplementary movies (11) and validation artificial cells; full data on requestartificial cells: known contraction amplitude and timing; real movies: none spatialone imagerequestdata on reasonable request; article CC BY; macro GPL-3.0askneeds a request
the data needs a request or a data use agreement
Mostly 1D contraction signals, not dense displacement Licence checked on 2026-10-10 at https://pmc.ncbi.nlm.nih.gov/articles/PMC5805275/; derived test pairs shareable: unclear; commercial use: unclear. Circ Res paper: 'datasets generated and analyzed ... are available from the corresponding authors on reasonable request'; movies shown as supplements fall under the article's CC BY only if they are part of it (unclear); repo holds only the macro and manualsynthetic contraction of a real frame with a known displacement field
NAOMi-based synthetic calcium imaging with scan motion (Shen 2024)simulated two-photon calcium movies (modified NAOMi) with motion artefacts, frame alignmenttraining and test sets (count unverified)known simulated motion and clean moviesboth imagesdirectnone statedaskwaits for a tool
simulated movies with known motion, no licence: unchanged; waits for point tracking
Synthetic; the NAOMi simulator itself can generate more with exact motion Licence checked on 2026-10-10 at https://zenodo.org/records/8248112; derived test pairs shareable: unclear; commercial use: unclear. Zenodo record is open access but has no licence field; it asks to cite Shen et al., PhotoniX 2024
Neurofindertwo-photon calcium imaging movies from several labs, frames of one session19 training + 9 test movies (per CaImAn paper comparison; unverified)neuron ROIs (varied quality per lab), no motion truthone imagedirectnone statedaskwaits for a tool
movies with ROIs but no motion truth, no licence: unchanged; waits for point tracking
Superseded by CaImAn annotations for consistency Licence checked on 2026-10-10 at https://github.com/codeneuro/neurofinder; derived test pairs shareable: unclear; commercial use: unclear. README lists datasets 'generously provided by' several labs, with no data licence; neurofinder.codeneuro.org timed outknown synthetic motion added to stable segments
NIST MIST reference stitching dataphase contrast and Cy5 tile grids of fixed stem cell colonies, tile-to-tile stitching3 plates x 5 overlaps (10-50%), 1,150 reference coloniesstage coordinates of each colony centred in the field of view (feedback loop, <1 um) and colony areasboth imagesdirect: three zips on isg.nist.govnone on the data page; NIST-made data is not subject to US copyright (17 U.S.C. 105)openleft out
Three plates (18.8, 20.1, 21.2 GB zips) of phase-contrast and Cy5 tile grids at 10 to 50% overlap; the reference is the stage position of 1,150 colonies centred by an automated feedback loop on thresholded images (NIST page). It scores where a stitched mosaic places colonies, a translation-only tiling problem: two overlapping tiles leave nothing for a registration method once the corpus has aligned them rigidly.
Physical stage reference, not pixel-exact; the standard stitching accuracy benchmark Licence checked on 2026-10-10 at https://www.nist.gov/open/copyright-fair-use-and-licensing-statements-srd-data-software-and-technical-series-publications; derived test pairs shareable: yes; commercial use: yes. the isg.nist.gov data page names no licence; NIST's policy covers NIST-created data (copyright may apply outside the US), acknowledge NIST
NP-Cloud SMLM drift correction data (Nat Commun 2025)single-molecule localisation microscopy frames/localisations over time with sample drift, frame-to-frame (or time-block) alignmentsimulated and experimental datasets (count unverified)simulated: known drift trajectory; experimental: fiducial or comparison methods (unverified)both imagesdirectCC BY-NC 4.0non-commercialwaits for a tool
localisation point clouds, non-commercial: unchanged; waits for mesh/point-cloud scoring
Localisation-level data (point clouds), not pixel images; registration of point sets over timeinject a known smooth drift into a stable localisation set
NucleiSky datasetsCross-scale fluorescence, live-to-fixed, 3D localisation10 datasetsBenchmark localisations; nuclei segmentationsdirect: Zenodo record 20721416CC BY 4.0 (Zenodo)openleft out
Zenodo (CC BY 4.0, 5.5 GB, eight zips): the 2D localisation benchmarks (datasets 1, 2, 4) take query crops out of the reference image itself, and the cross-scale (5, 6), live-to-fixed (7) and 3D (10) sets are one example each, located by NucleiSky with no independent truth. Similarity placements of crops leave nothing deformable to score.
Nuclei centroids matched by a program and verified by a person give dense points between live and fixed images.
OCT vs Bessel two-photon angiograms of the same capillaries (DANDI 001537)OCT angiography vs Bessel-beam two-photon angiograms of mouse cortex, same day, same animal20-minute series per modality (animal count unverified)nonedirectCC BY 4.0opentruth to make
Bessel two-photon projections vs OCT angiograms of the same capillaries, 377 GB, no truth; bifurcations found by a program and checked by a person; animal count unverified
made to compare capillary stalls between modalitiescapillary bifurcations found by a program then verified by a person in both modalities
Open-ST 3D metastatic lymph node and mouse head (Schott 2024 Cell)sequencing-based spatial transcriptomics and H&E of the same fresh-frozen section; 19 serial Open-ST sections plus 11 H&E-only sections of one HNSCC lymph node metastasis over 350 um; also E13 mouse head, adult mouse hippocampus19 same-section ST/H&E pairs; 30 serial sections for 3Dauthors' alignment (openst pairwise modality alignment with circular capture-area marks; STIM for 3D); no landmark truthone imagedirectGEO GSE251926 (public); paper CC BYopentruth to make
same-section ST/H&E and 30 serial sections of one lymph node, the authors' alignment only; one specimen for the serial part
capture-area fiducial marks visible in H&E images help modality alignmentsame section: nuclei from H&E vs transcript density peaks; serial sections: vessel and follicle landmarks by two people on 10 pairs
Oral cancer LBC cytology, brightfield and fluorescence of the same Pap-stained slides (Lu 2024, Uppsala)brightfield (11 z-levels) and 4-channel fluorescence (5 z-levels) whole-slide scans of the same Pap-stained liquid-based oral cytology slides, two scanners19 patients; 766,565 aligned cell patch pairs announcedglobal rigid (globalign) plus per-nucleus translation by the authorsboth imageshiddennot released ('in preparation' in the preprint); research use announcedunknownleft out
its images or its truth are not released
rare same-slide cytology multimodal pair with a known scale factor 1.472 between scanners; whole slides would be the useful partnuclei centroids on both scans (same cells), matched after the global fit, person verifies; scanner stage drift is the real deformation
Oral squamous cell carcinoma MALDI MSI with H&E and pathologist labels (Zenodo 7377802)MALDI mass spectrometry imaging of OSCC cryosections with an H&E image and pathologist ground-truth map2 labelled sections from 5 patients (45,738 spectra raw)pathologist region labels drawn on the optical scan; MSI-H&E registration not describedone imagedirectCC BY 4.0openleft out
two labelled sections, labels on the optical scan only: too few pairs
whether H&E is the same or an adjacent section is not statedtumour/healthy boundary traced on H&E and on an MSI cluster map by a person; ablation pattern if visible
ORION-CRC (Lin, Chen, Campton 2023)Same tissue section: 16-18 plex Orion immunofluorescence (Hoechst nuclear channel) then H&E, colorectal cancer, same subject40 whole-slide specimens (CRC01-CRC40), each H&E vs IF of the same sectionNone for registration; H&E released already registered to the IF frame by the authors (file '-registered.ome.tif'), plus cell masks and tables in IF spaceone imagedirectCC BY 4.0 (HTAN open-access imaging data); Zenodo index record MITopenleft out
same Lin 2023 Orion data as the 'Orion same-section H&E and 18-plex IF, colorectal cancer (Lin 2023)' row (2D whole slides, IF about 113 GB each)
Raw unregistered H&E scans not obviously listed beside the registered ones (unverified); MIPHEI-ViT (zenodo 15340874) redistributes aligned tiles Licence checked on 2026-10-10 at https://humantumoratlas.org/data-access; derived test pairs shareable: yes; commercial use: yes. Zenodo 10.5281/zenodo.7637988 (MIT) only indexes files on s3://lin-2023-orion-crc; the images are also released through HTAN, whose open-access imaging is CC BY 4.0; same data as the Orion HTAN rowSame section, so Hoechst vs hematoxylin nuclei give dense truth: use the authors' registered H&E as the known warp, apply a known smooth deformation, or have two people place 20 nuclear landmarks on 10 pairs
Paired H&E and IHC whole-slide images, mouse colitis (Dryad 2023)H&E vs IHC whole-slide images of mouse colon (same or serial sections, unverified)paired WSIs (count unverified); IHC parts in Dryad ffbg79d0q and wm37pvmt0nonedirectCC0opentruth to make
H&E and IHC whole slides of mouse colon (CC0, 147 GB for the H&E part), serial or same section unverified, no truth
made for computational IHC-H&E mappingtwo people mark 15 crypt and vessel landmarks per pair on 10 pairs
PANDA (prostate biopsy grading)prostate biopsy whole-slide imagesnoneGleason grades, some masksone imageaccountKaggle competition rulesnon-commercialneeds an account
the download needs an account
Fails pairs: one image per subject, no repeat or second modality. One stain per biopsy. Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Foundation models review 2026).One stain per slide: a known warp on one slide, or pair slides of the same biopsy if any were scanned twice.
PARS virtual vs chemically stained whole slides (BioImage Archive S-BIAD2232)virtual stains from label-free dual-excitation PARS microscopy vs the chemically stained same sectionpaired WSIs (count unverified)none stateddirectCC BY 4.0openleft out
only network-made virtual stains (RegGAN) against the chemical stain; the label-free PARS images are not released
virtual images were generated by RegGAN, which itself registerssame section: nuclei matching by a program checked by a person; the raw PARS images would be the harder multimodal input
PASTE preprocessed spatial transcriptomics slices (SCC, Stahl breast, HER2 BC, DLPFC)Spot-based spatial transcriptomics of consecutive sections: cutaneous squamous cell carcinoma (Ji 2020; patients 2, 5, 9, 10, 3 replicates each), breast cancer layers (Stahl 2016), HER2+ breast cancer, DLPFCAdjacent replicate pairs per patient (SCC ~8 pairs; breast 3-4 layers)None beyond DLPFC layer labels; alignment judged by expression similarityone imagedirectCC BY 4.0openwaits for a tool
spot coordinates and counts, no images: unchanged; waits for mesh/point-cloud scoring
Count matrices and spot coordinates only; H&E images must come from the original deposits Used in 1 of the 2D and microscopy registration papers read (GPSA 2023).Fetch the original H&E images and have two people place 15 landmarks per adjacent pair
Pediatric Crohn's multi-stain restained WSIs (Ounissi 2024)Same tissue section H&E then restained with one of 8 stains (AE1/AE3, CD117, CD15, CD163, CD3, CD8, D2-40, Giemsa), pediatric Crohn's biopsies240 H&E/restain WSI pairs (30 per stain), 0.22 um/pxNone; pairs described as 'perfectly matched' identical sectionsone imagedirect: Zenodo record 16751426, one zip readable by rangesCC BY 4.0 (Zenodo)openleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): open, but not among the six minor sets kept. It was proposed as: Zenodo (CC BY 4.0, one 4.8 GB zip readable by ranges): about 5,170 paired tiles of 1024x512 px, H&E and one of eight IHC or Giemsa stains of the identical section side by side (512x512 each, 0.22 um/px), from about 160 slide pairs (11 to 30 per stain, not the 480 WSIs the description counts) of 59 pediatric Crohn's patients. Aligned by construction, so the truth would be a known warp (synthetic), ranked apart.
Restained same section: residual non-rigid deformation from restaining and rescanningRestained same section, so register nuclei and have a person verify 20 landmarks on 10 pairs; or apply a known warp after alignment for exact truth
PESO (prostate epithelium, H&E restained with P63/CK8-18 IHC)Prostatectomy H&E slides restained with P63 and CK8/18 IHC (same section)None directly: 62 training H&E WSIs released with IHC colour-deconvolution masks, but not the IHC WSIsIHC-derived epithelium masks (already mapped to H&E by the authors); 40 test regions with manual masksone imagedirectCC BY-NC-SA 4.0non-commercialleft out
only the H&E slides and IHC-derived masks are released, not the restained IHC slides: no second image
Fails as a pair dataset unless the IHC WSIs are obtained from the authorsAsk Radboud for the restained IHC WSIs; same section, so a known warp gives exact truth
PLISMSame tissue on 7 scanners + smartphones, 13 H&E conditions4,454 phone images, 3,417 WSIsPrecisely aligned patches (per WSI review)both imagesdirect: figshare+ (no account), per-file downloads; PLISM-wsi tiles in one tar.gz per stain and device (1.1-1.6 GB), original WSIs one file each (1-2.6 GB)CC BY 4.0 (figshare+ records 23614422, 24988074 and 23590791)openleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): open, but not among the six minor sets kept. It was proposed as: PLISM (three figshare+ records, all CC BY 4.0): one tissue microarray of 46 human tissues, 13 serial sections each stained under another H&E condition and scanned by 7 slide scanners (91 original WSIs, 181 GB), plus 6 smartphones through a microscope. The alignment is the provider's own registration (VALIS rigid and non-rigid, then AKAZE rigid fits; reported TRE 43 µm between serial sections, 325 hand-refined points on 65 patches only to check the WSI-smartphone step, not released), so the corpus would make the truth, a known warp on same-section cross-scanner pairs, and rank it apart.
Cross-device alignment status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (WSI registration review 2025). Licence checked on 2026-10-10 at https://doi.org/10.25452/figshare.plus.23614422.v1; derived test pairs shareable: yes; commercial use: yes. Figshare+ records for the WSI, original WSI and smartphone subsets are CC BY 4.0 (DataCite); Owkin HF tiles also CC BY 4.0
Porcine kidney ex vivo MRI and cleared-tissue microscopy (UQ)16.4T MRI (T2, T2*, DWI, QSM) and confocal microscopy of cleared kidney cortex, same sample6 sampleslandmark-based co-registration to locate single glomeruli (landmarks release unverified); MRI masksdirectUQ General Usage Terms and Conditions for data publishing in eSpaceasktruth to make
3D 16.4T MRI and cleared-tissue confocal of the same six samples; the landmarks used to co-register them are not among the four released folders; UQ terms variant not readable (record page 403)
30 µm MRI; glomeruli are natural point landmarks. Licence checked on 2026-10-10 at https://doi.org/10.48610/6c30dfb; derived test pairs shareable: unclear; commercial use: unclear. DataCite rights name the UQ general terms; UQ offers four variants (attribution, non-commercial, share-alike, data sharing agreement) and the record page (403 here) would say which; terms URL now 404glomerulus centres found by a program in both, verified by a person
Princeton Mouse Atlas (PMA) and aligned transsynaptic tracing brains (Pisano 2021)iDISCO+ light-sheet whole mouse brains registered to a light-sheet template (PMA) with CCF labels; between subjectsTemplate plus HSV/PRV injected brains aligned to PMA (counts not checked)69 landmark points by 11 blinded users (median 93 um after registration, repeat-placement floor 49 um); release of the points unverifiedboth imagesdirectCC BY-NC-ND 4.0non-commercialneeds a request
only the atlas and the aligned injection sites are public; the raw light-sheet brains are shared by the lead contact on request, and the landmark points are not released
Aligned injection data at brainmaps.princeton.edu/2021/05/pisano_viral_tracing_injections/ Licence checked on 2026-10-10 at https://brainmaps.princeton.edu/2020/09/princeton-mouse-brain-atlas-links/; derived test pairs shareable: no; commercial use: no. 'These files are shared under the CC BY-NC-ND license'; no-derivatives means resampled test pairs cannot be shared without permission; the tracing brains of Pisano 2021 were not checked separatelyRe-place the landmark protocol (2 people, 20 points per brain) on 10 raw brains if the original points are not released
RCC H&E slides digitised on multiple scanners (BioImage Archive S-BIAD1343)same H&E glass slide scanned on Zeiss, Hamamatsu, Philips, Leica and other scanners100 cases, several scans eachnone; same physical slideBioImage Archive, direct per fileCC BY 4.0 (BioImage Archive study)openleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): open, but not among the six minor sets kept. It was proposed as: 100 ccRCC H&E slides of the SCOTRRCC cohort scanned on six scanners (two Zeiss Axio Z1, Hamamatsu NanoZoomer 2.0-HT and S60, Philips UFS, Leica AT2; 0.22 to 0.50 µm/px): 599 whole slides, 586 GB, median 0.83 GB, in czi, ndpi, svs and Philips iSyntax (BioImage Archive). No annotation: the same glass slide on two scanners is rigid apart from stitching, so a known warp (synthetic) on gigapixel pairs is the only truth; the Wilm set (registration 1144) gives the same setting smaller.
near-rigid problem with scanner colour and resolution change20 nucleus landmarks per scan pair by two people on 10 cases; a fine automatic alignment then verified
REACTIVAS renal biopsy H&E, PAS and multiplex IF pairsRenal biopsy sections, brightfield H&E/PAS vs multiplex immunofluorescence; same sample11 paired samplesNot statedrequestprivate clinical datasetaskneeds a request
the data needs a request or a data use agreement
Origin and access not described in the excerpt read (unverified) Licence checked on 2026-10-10 at https://arxiv.org/abs/2511.03826; derived test pairs shareable: no; commercial use: no. CORE paper evaluates 'four public datasets and two private clinical datasets'; REACTIVAS comes from clinical trial NCT04916704 and is one of the private onesIf IF and brightfield are on the same section, nuclei (DAPI vs haematoxylin) give dense truth after a person checks 20 points
Registered angular mineral segmentation, polarized-light thin sectionspolarized-light microscopy of rock thin sections at several stage rotations, same section97 thin sections, rotation sequencesSIFT registration by the authors; 9,231 shared masks reviewed by geologistsZenodo 21704195, direct, zip readable by ranges (the DOI itself resolves to IEEE DataPort, which needs a login); an identical figshare copyCC BY 4.0 (Zenodo)openleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): not biomedical. It was proposed as: 97 rock thin sections (Nanjing University set), eight polarised-light JPEGs per section at stage rotations, already registered by the authors' SIFT step so that one mask (eight mineral classes, SAM then corrected by geologists) serves the eight views (Zenodo record, zip listing). As delivered the views are aligned and nothing is left to register: a known warp between two angles, whose colours change with extinction, would make a synthetic non-medical 2D set.
also on Zenodo (record 21704195); colour changes strongly with rotationtwo people mark 10 grain-corner points across the rotations of 10 sequences
Reinforced optical cage SMLM bead and STORM datatime series of fluorescent beads (homebuilt and commercial microscopes) and STORM of HeLa cells, lateral and axial drift5 + 5 bead series, 2 STORM seriesbead sub-pixel localisations per frame (physical fiducials)both imagesdirectCC BY 4.0openwaits for a tool
bead and STORM frame series with per-frame bead positions: a drift (tracking) problem, unchanged; waits for point tracking
Beads give exact per-frame drift on real images: clean rigid-drift truth
RnR-ExM (Robust Non-rigid Registration for Expansion Microscopy, ISBI 2023)3D expansion microscopy, multi-round imaging of same sample (zebrafish brain, mouse cortex, C. elegans)4 train pairs per species + validation + test pairs (pair5-7) on Zenodo 7444100/8418961/8419002Segmentation labels of fixed and moving volumes (organisers) used for Dice/HD95both imagesdirectCC BY 4.0 (Zenodo)opentruth to make
3D ExM volume pairs on Zenodo (CC BY 4.0), but the organisers keep the evaluation segmentations hidden; nuclei and vessel landmarks by two people would make it
Strong z-offsets, repetitive texture, gel deformation; voxel 0.1625x0.1625x0.4 um
Same-tissue Xenium, COMET and H&E with alignment matrices (UnumLocalia)Xenium transcripts and DAPI vs COMET multiplex IF vs H&E on the same core4 cores (HCC tumour, non-tumour liver, tonsil, hepatocellular adenoma)alignment matrices from COMET and H&E to Xenium (authors)both imagesdirectCC BY 4.0openleft out
same Zenodo record (21713660) as the 'UnumLocalia same-section Xenium, COMET and H&E cores (Duchini 2026)' row
matrices are a made alignment, probably affinenuclei matched DAPI to haematoxylin by a program, 20 per core verified by a person
SegPath (restaining-based annotation, Komura 2023)H&E sections destained and restained by immunofluorescence (8 cell-type antigens), many cancer typesNone usable: only H&E tiles and binary masks derived from IF are released (one Zenodo record per antigen)Masks registered to H&E by the authorsone imagedirectCC BY-NC-SA 4.0non-commercialleft out
only H&E tiles and IF-derived masks are released, not the IF images: no second image
Fails: the IF images themselves are not released, so there is no image pair to registerMask vs H&E could test segmentation-guided registration only after a synthetic warp; ask the authors for the raw IF
Seminoma H&E and Ki67 neighbouring-section pairs (Zenodo 11218961)H&E vs Ki67 IHC on neighbouring sections of testicular seminoma77 pairspairs registered and aligned by the authors (method unverified)directCC BY 4.0opentruth to make
77 neighbouring-section H&E/Ki67 pairs given already affine-aligned (PNG, about 35k px square, 197 GB, CC BY 4.0), raw scans not released; people's points would measure the non-rigid residual
only the registered result may be given; check for the raw imagestwo people place 15 nucleus or gland landmarks per pair on 10 pairs
Serial Section Transmission Electron Microscopy (Zenodo 13887963)serial section EM images, section-to-section alignmentone series (size 2.3 GB)none statedone imagedirect: Zenodo, two zips (0.54 and 1.75 GB) whose members read by HTTP rangesCC BY-SA 4.0 on the Zenodo recordopenleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): open, but not among the six minor sets kept. It was proposed as: Two zips of 4096x3488 8-bit TEM images from Mount Sinai (AMT camera, 681 pixels a micrometre, about 1.5 nm a pixel): sample 233L as 15 numbered sections x 10 lettered images (150) and sample 36L as 34, with low-magnification maps edited in Photoshop. The record says only 'Serial electron microcopy images' (no tissue, section thickness or order), carries no labels and the sections are not aligned, so any truth is the corpus's (a known warp of one section, or points placed by people on consecutive sections); two samples cannot give ten test and ten development pairs from different subjects.
Minimal descriptionknown warp per section, or label tracing of 20 neurites across sections
Simulated partial-overlap DLPFC slice pairs (Hu et al. 2024 benchmark)Synthetic second slice made from one DLPFC Visium slice with 20-100% overlap and expression noise; same slice, known spot correspondence5 overlap ratios x noise levels per reference sliceExact spot-to-spot correspondence by construction (simulation by the benchmark authors)both imagesdirectnone statedaskwaits for a tool
simulated spot-level pairs without images: unchanged; waits for mesh/point-cloud scoring
Whether the simulated pairs themselves are deposited (vs the simulation code) is unverified; annotated DLPFC on Zenodo 10698880 Licence checked on 2026-10-10 at https://benchmarkst-reproducibility.readthedocs.io/en/latest/Data%20availability.html; derived test pairs shareable: unclear; commercial use: unclear. data availability page lists source datasets (DLPFC from spatialLIBD) and links with no licence
SLICE-1 and SLICE-2 Tribolium light-sheet embryogenesismulti-view light-sheet time-lapse of beetle embryos (four directions), raw views and fusions50 datasets (SLICE-1, 2.7 TB) plus 16 isotropic gastrulation datasets (SLICE-2)none for registration; fusions made by the authorsone imagedirectunverified (Zenodo)unknowntruth to make
3D multi-view light-sheet time-lapses (Zenodo, CC BY 4.0, about 70 GB a dataset), no registration truth; nuclei matched across views would make it
Sci Data 2025 and 2026. Very large time-lapse volumes.nuclei detected in overlapping views matched by a program and verified; or fluorescent beads in the mount as fiducials
Slide-seqV2 mouse hippocampus pucks (Stickels 2021)Slide-seqV2 bead-level expression maps, mouse hippocampus (also embryo, olfactory bulb, cortex pucks); two animals, between subjects2 hippocampus pucks used as a pair (Puck_191204_01, Puck_200115_08); 14 pucks in the studyCell-type labels per bead; GPSA placed 3 landmarks (dentate gyrus ends, CA3 edge); Hu 2024 drew CA1-3 and DG regionsboth imagesaccountSingle Cell Portal terms: public study data 'made available for unrestricted public view, redistribution and reuse'openneeds an account
the download needs an account
Point clouds, not images; field of view differs between pucks, needs coarse manual rotation first Licence checked on 2026-10-10 at https://singlecell.broadinstitute.org/single_cell/terms_of_service; derived test pairs shareable: yes; commercial use: yes. no named licence on SCP815; the SCP terms state public-study data are open for redistribution and reuse; downloads need a Google sign-inTwo people place 15 landmarks on hippocampal layer ends in both pucks; or map both to Allen CCF and compare region labels, a weaker check
SNEMI3D (ssSEM mouse cortex)serial section SEM stack of mouse cortex (Kasthuri), already aligned; section-to-section alignment by construction100 training + 100 test sectionsdense neuron labels on every training sectionboth imagesdirect: Zenodo record 7142003, no accountCC BY 4.0 (Zenodo record)openin the corpuspriority 2
Zenodo 7142003 (CC BY 4.0, 0.19 GB) holds snemi.zip: a 100-section serial-section SEM training stack of mouse cortex (1024x1024) with every neurite labelled by an expert neuro-anatomist (Daniel Berger), and a 100-section test stack without labels, both cut from the Kasthuri et al. (Cell 2015) volume. The stack is already aligned, so pairs come from known smooth warps drawn by the corpus (synthetic), and all sections come from one animal: test pairs from the test stack and development pairs from the labelled stack at best.
Pre-aligned: registration problems must be made by applying known warps Licence checked on 2026-10-10 at https://doi.org/10.5281/zenodo.7142003; derived test pairs shareable: yes; commercial use: yes. training and test data on Zenodo as linked from the challenge pageapply known smooth in-plane warps per section: exact truth plus labels
Spinal cord diffusion tensor microscopy with histology (Flint 2016)15.6 µm DTI microscopy and histology of human, pig and rat spinal cord sections9 datasetscorresponding histology, no registration truthone imagedirectCC BY 4.0openwaits for a tool
2D histology against DTI microscopy volumes: unchanged; waits for 2D/3D pose error
Hosting not checked. Licence checked on 2026-10-10 at https://data.mendeley.com/datasets/h3wkngnptj/1; derived test pairs shareable: yes; commercial use: yes. Mendeley Data record (DataCite rights CC BY 4.0), cited by the Data in Brief articletwo people mark grey-white matter boundary points in both
StainBridge renal biopsy serial WSI cohort (Weill Cornell)Consecutive renal biopsy sections, 4 structural stains (H&E, PAS, PAMS, trichrome) and 10 IHC; same biopsy315 consecutive-section pairs from 23 cases, 338 WSIs; 272 pairs landmarked1,468 manual glomerular landmark correspondences (glomerulus polygons in QuPath)both imageshiddenNot statedunknownleft out
its images or its truth are not released
Private as far as stated; 3D reconstruction use case, many IHC-IHC pairsAsk the authors; for a public analogue, two people mark glomerulus centres on KPMP or ANHIR kidney sets
Stars2Cells cross-session neuron tracking data (Zenodo 20938743)calcium imaging footprints of the same neurons across sessions (striatum) plus a synthetic benchmark1,262 synthetic paired runs; real sessions from one studyground-truth identity tables for the synthetic runs; method assignments for real datadirectCC BY 4.0openwaits for a tool
cell footprints and centroid sets, raw videos on request: unchanged; waits for mesh/point-cloud scoring
footprints, not raw images, for the real datatwo people verify 100 cross-session matches in the real data
STHELAR (Xenium spatial transcriptomics with H&E)Xenium DAPI vs H&E of the same slide, 16 tissue types31 whole-slide imagesco-registered DAPI and H&E images provided (made by the authors)both imagesdirectCC BY 4.0opentruth to make
31 Xenium DAPI/H&E sections of 16 tissues co-registered by 10x/the authors (BioImage Archive, CC BY 4.0, direct): same task as the 10x Xenium row, one copy only; nuclei pairs checked by a person would make it main-board
registration used as given; unregistered originals unverifiedcheck the given alignment with 20 nuclei per slide verified by a person
Targeted volume CLEM of marine microorganisms (EMPIAR-11399)fluorescence light microscopy vs aligned FIB-SEM slices of environmental microorganismscount unverifiednone stateddirectCC0 (EMPIAR default, unverified)openleft out
EMPIAR holds only the aligned FIB-SEM volume and names no light-microscopy deposit: no second image
matching LM data location unverifiedorganelle landmarks by two people
TCGA frozen vs FFPE slides of the same patient (IDC TCGA collections)Frozen-section H&E (tissue slides) and FFPE diagnostic H&E slides of the same TCGA patients, 33 cancer typesThousands of frozen-FFPE slide pairs by patient, usually from different blocksNoneone imagedirectCC BY 3.0 for most TCGA slide collections in IDC (per IDC Zenodo pages)openleft out
frozen and FFPE slides of a patient usually come from different tissue pieces, so the pairs do not show the same tissue; no truth
Mostly fails for registration: frozen and FFPE slides generally come from different tissue pieces, so only a minority are true correspondences; frozen-to-FFPE translation papers use them unpairedScreen for pairs from the same tissue piece by shape matching, then two people place 15 landmarks on 10 confirmed pairs
TIGER (TILs in breast cancer) challenge dataH&E whole-slide images of HER2+ and triple-negative breast cancer with tissue, cell and TIL annotationsNone: H&E only, one slide per caseSegmentation and lymphocyte detections on single imagesone imagedirectCC BY-NC 4.0 (annotations, RUMC and JB slides); TCGA-BRCA slides under TCGA termsnon-commercialleft out
H&E only, one slide per case: no second image
Fails: no second stain or second image to register; listed because it was named as a lead Licence checked on 2026-10-10 at https://tiger.grand-challenge.org/Data/; derived test pairs shareable: yes; commercial use: no. Data page 'Data license' section; public training set on AWS Open Data (s3://tiger-training/, no account)Only synthetic warps of single images; not a real registration benchmark
Time-lapse 3D confocal of spruce wood enzymatic deconstruction (Data in Brief 2025)time-lapse 3D confocal of wood cell walls deforming under enzymatic hydrolysis, same sample over hourshydrolysis and control time-lapsescell segmentations tracked over time (6,699 and 6,505 tracked cells)both imagesdirectunverifiedunknowntruth to make
3D confocal time-lapses (25 hourly stacks, 6 samples) with cell labels per time point, but the identities come from the authors' registration (HydroTrack), no manual check stated; a person audits the tracks
Non-medical biology; tracked cell labels give correspondence under large deformation.a person audits the cell tracks on 3 time pairs; cell centroids then serve as landmarks
TNBC H&E restained with PHH3 (Tellez 2018, Radboudumc)Same breast cancer section scanned as H&E, destained, restained with PHH3 and rescanned; same section18 WSI pairsMitosis detections registered from PHH3 to H&E (CNN plus manual labels); no landmark setone imagerequestnot statedunknownneeds a request
the data needs a request or a data use agreement
No public release found (unverified); restained = near-rigid, ideal for exact multimodal truth Licence checked on 2026-10-10 at https://arxiv.org/abs/1808.05896; derived test pairs shareable: unclear; commercial use: unclear. paper describes the TNBC H&E/PHH3 restained pairs but never says they are released; no data record found; ask Radboudumc (DIAG)Restained sections are aligned by construction up to scan shift: a known warp gives exact H&E-to-IHC truth
Tonsil and lung cancer t-CyCIF raw images (Rashid 2019 Sci Data)t-CyCIF of reactive tonsil and 3 lung cancer specimens, 11-12 cycles, Hoechst (DAPI) every cycle; plus 8 antibody-validation sections4 specimens x 11-12 cycles of raw rcpnl tiles, plus stitched/registered OME-TIFFNone; ASHLAR registrationone imageaccountCC BY-SA 4.0 (data, per the article)openneeds an account
the download needs an account
Raw per-cycle RareCyte tiles released, a rare feature; Synapse account neededDAPI in each cycle: matched nuclei verified by a person; or compare cycle 1 vs re-imaged cycle 1 (DATASET-2 imaged twice) as a near-identity test
TP53-Precursor-Lesions (TCIA 2026)H&E, p53 and Ki-67 sections of fallopian tube from the same blocks (20x camera photos)sections of 78 subjects in the collection (study of 242 women)none for alignment; pathologist reviewone imagedirectunverified (TCIA collection licence)unknownleft out
camera photos of chosen fields, not slides: the fields of two stains need not show the same tissue
JPEG photos of fields, not whole slides; fields may not match across stains.two people place 10 landmarks (gland corners, cell clusters) per stain pair on 20 pairs
Track2p longitudinal calcium imaging in developing mouse cortex (Zenodo 17091226)two-photon calcium imaging of barrel cortex in mouse pups across one postnatal week, same cellsseveral mice, daily sessionsground-truth data for evaluating cell trackingboth imagesdirectCC BY 4.0openneeds a request
the README lists traces, behaviour and tracking ground truth for 6 mice but no images; raw two-photon movies only on request
the brain grows between sessions, a real deformation
Tracking neurons in a moving and deforming brain (Nguyen 2017)Whole-brain imaging of freely moving C. elegans; neuron tracking through large deformationFrame pairs in recordings (counts not checked)Manual neuron annotations (unverified)both imagesaccountIEEE DataPort open access; no licence shownaskneeds an account
the download needs an account
Details from citations only; page not opened Licence checked on 2026-10-10 at https://ieee-dataport.org/open-access/tracking-neurons-moving-and-deforming-brain-dataset; derived test pairs shareable: unclear; commercial use: unclear. record DOI 10.21227/H2901H has no rights in DataCite and the page names none; download needs a free IEEE accountTwo people label 20 neurons across 50 frames
TumorQuantAI colon cancer CD3/CD8/CK20 serial sectionsSerial sections of colorectal cancer stained CD3, CD8 and CK20 IHC, same subject9 complete triplets (27 consecutive-stain pairs), 30 WSIs at 0.26 um/pxAutomatic registration QC (numeric and visual) by the authors; no landmarksone imagedirect'Copyright (C) 2026 The Authors'; no open licenceaskneeds a request
copyright reserved ('research and technical review only'), proprietary Motic .mds slides, 9 patients: ask the authors
Proprietary Motic MDS format may need conversion; posted 2026-08 Licence checked on 2026-10-10 at https://zenodo.org/records/22177196; derived test pairs shareable: unclear; commercial use: unclear. record text: 'Public visibility does not waive copyright, imply a separate open-content license'; research and technical review onlyTwo people place 20 landmarks (crypts, vessels) on each of the 27 pairs
Twocan-954 paired IF and IMC imagesimmunofluorescence and imaging mass cytometry of the same co-stained slide; cell lines and mouse lung954 pairs (864 mouse lung, 90 cell lines); qc_pass with transformstransformation found by Bayesian-optimised registration, QC by the authorsboth imagesdirectCC BY-SA 4.0opentruth to make
954 same-slide IF/IMC pairs with the authors' Bayesian-optimised transforms (CC BY-SA 4.0, 13.7 GB); IMC after IF on one slide is likely near affine, so little room unless nuclei pairs show more
same slide, so near-rigid; truth is the authors' best registration, not independentnuclei detected on IF DAPI and IMC DNA, matched, person verifies 20 on a test subset
UCLA transplant virtual staining example pairs (Li 2024, lung and heart)4-channel autofluorescence of label-free lung and heart biopsies and the same fields after H&E, Masson trichrome, EVGexample fields only (number unverified)fields registered by the authors (affine then learned)both imagesdirect: Zenodo, one 4.7 GB RAR (inputs stored uncompressed, targets compressed; not readable by ranges with standard tools)CC BY 4.0 on the Zenodo recordopenleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): open, but not among the six minor sets kept. It was proposed as: The 4.7 GB RAR (listed through its RAR5 headers) holds five model folders and five example folders (heart H&E and MT, lung H&E, MT and EVG), each with 10 or 11 tiles of one region ('Region_1_tv-XX_th-YY'): a 57 MB four-channel autofluorescence .mat input, the network output and the histochemically stained target of the same field. About 54 aligned pairs from five tissue regions: the only truth is the corpus's known warp on the authors' alignment, and five regions cannot give ten test and ten development subjects.
small example set inside a code archive; the full >3 TB data is not publicknown warp on the aligned stained image
UniStain multi-organ autofluorescence and H&E pairs (HKUST)label-free autofluorescence and H&E of the same FFPE section; lung, liver, kidney, breast6,800 patch pairs from 30 sectionspairs pre-registered by the authorsboth imagesdirectresearch use only (README)non-commercialleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): its licence is not shown to allow commercial use (research use only (README)). It was proposed as: same-section autofluorescence/H&E patches pre-registered by the authors, 30 sections of 4 organs: a known warp; research use only
Google Drive download; patches are already aligned, so whole-slide originals would be needed for registration tests (availability unverified)same section: apply a known smooth warp to the aligned H&E for exact multimodal truth
Unsupervised multimodal microscopy registration datasetsMultimodal microscopy of same field (renal cancer tissue, HeLa, skin)3 datasetsUnverifieddirectCC BY 4.0opentruth to make
screening vs laser-microdissection microscope pairs (119 HeLa, 107 renal, 79 skin, per the SuperCUT paper) with at least three people's point pairs each, used for a similarity (whether the point files are in the zips is not stated); drying bends the sample, so more points per pair would make a deformable truth; CC BY 4.0
Pairs show the same field and are aligned by construction: a known warp on one image gives exact multimodal truth (check the alignment first).
UnumLocalia same-section Xenium, COMET and H&E cores (Duchini 2026)Xenium transcripts, Lunaphore COMET 40-plex IF and H&E on the same FFPE section; HCC tumour, non-tumour liver, tonsil, adenoma4 cores, three modalities eachalignment matrices from the authors (COMET and H&E to Xenium)both imagesdirectCC BY 4.0openleft out
four cores: too few pairs for a test set
small but clean; authors' matrices are their own registration, not independent truthnuclei matched across Xenium DAPI, COMET DAPI and H&E; person verifies 20 per pair
Vizgen MERFISH Mouse Brain Receptor MapMERFISH with DAPI mosaics, adult mouse brain coronal sections, 3 positions x 3 biological replicates9 sections: replicate pairs at matched positions (between-animal) and section-to-atlasNone; cell segmentation polygonsone imageaccountno named licence; access form with user indemnity; 'Vizgen Data Release V1.0'askneeds an account
the download needs an account
Between-animal pairs, matched position not exact status change: landmarks now published in the STalign supplement, so the entry has a point truth (between-animal, coarse) Licence checked on 2026-10-10 at https://info.vizgen.com/mouse-brain-map; derived test pairs shareable: unclear; commercial use: unclear. page says 'Copyright 2026 Vizgen. All Rights Reserved'; launch announcement says participants may use the data as they choose, but no licence text; also on Google Cloud Marketplace (gcp-public-data-vizgen)Register each section to Allen CCF and compare region labels; or two people place 20 landmarks per replicate pair
Volume correlative light, X-ray and electron microscopy of cremaster endothelium (EMPIAR-11958)X-ray micro-CT of a resin block vs SBF-SEM of the targeted endothelial cells, mouse cremaster1 sample, several imagesetsnone stateddirectCC0 (EMPIAR default, unverified)openleft out
one sample (micro-CT of the block against one 80 GB SBF-SEM stack), no correspondences; chunked reading now exists, but one pair is a demo
found through EBI search for EMPIARvessel branch points in micro-CT and SBF-SEM marked by two people
Warpy WSI registration test (re-stained section)Mouse duodenum fluorescence then DAB re-stain of same section1 pairLandmarks per WSI review (unverified)directCC BY 4.0openleft out
one restained pair: a demo, not a test set
status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (WSI registration review 2025).Re-stained section: two people place 20 landmarks (nuclei, crypt features) on fluorescence and DAB; one pair, a demo case only.
Whole-brain 50-plex iterative IF rat brain (Maric 2021 Nat Commun)Iterative 10-plex immunofluorescence rounds (5 rounds, 50 markers) on whole rat brain sections, DAPI in each round, normal and injured brainRound-to-round pairs per section (DAPI to DAPI), whole-slice mosaics ~43k x 29k pxNone; authors' ORB+RANSAC affine on DAPI, claimed sub-pixelone imagedirectCC BY 4.0opentruth to make
5 IF rounds with DAPI each on whole rat brain slices (CC BY 4.0, 195 GB), the authors' affine only; matched nuclei checked by a person, though rounds may differ little beyond affine
Original acquired images and processed results both in the archive (unregistered rounds separated: unverified)DAPI repeats every round: matched nuclei verified by a person give near-exact truth; brain tissue deforms little between rounds
Wu, Temko, Maliga 2022 lung adenocarcinoma CyCIF and adjacent H&E with matched ROIsCyCIF 20+ plex and H&E of adjacent FFPE sections, 12 lung adenocarcinomas12 section pairs20-60 histologically defined regions per tumour with X/Y coordinates in both CyCIF and H&E (N_ROI.csv), by the authorsboth imagesdirectunverified (Harvard Tissue Atlas data, Synapse/AWS)unknownwaits for a toolpriority 1
waits for a tool: the H&E lies 81 to 126 degrees from the CyCIF on most patients and the corpus has no 2D rotation search for the start (elastix-rigid-oriented is 3D only). Region centres scatter a median 16 to 86 um per patient after an affine fit, so it would be ranked apart; no licence stated (ask). Reading is ready: CyCIF pyramid by ranges, H&E Olympus ETS tiles (JPEG, 0.346847 um).
ROI centres are region-level correspondences (coarse), not cell-level; raw tiles also allow tile stitching testsuse ROI centre pairs as coarse landmarks; add 20 nuclei-level landmarks per pair by two people
Xenium and MACSima on the same section (BioImage Archive S-BIAD3678)Xenium spatial RNA vs MACSima cyclic protein imaging of the same sectioncount unverifiednone released; registration code on GitHub (scOpenLab/MACSimaXenium)directCC0openleft out
one TMA core: too few pairs for a test set
found through the BioImage Archive APIDAPI in both: nuclei correspondences found by a program, checked by a person
Xenium FFPE human breast cancer with post-Xenium H&E and adjacent Visium (Janesick 2023)Same section: Xenium DAPI morphology image, then H&E and IF (CD20, HER2) on that section; adjacent section: Visium CytAssist with its own H&E; human breast cancer2 Xenium replicates: DAPI vs same-section H&E; Xenium vs adjacent Visium H&EKeypoint registration made in Xenium Explorer by 10x (companion notebook); no independent landmarksone imagedirectCC BY 4.0 (10x dataset pages); companion code under a 10x-restricted MIT variantopenleft out
same Janesick 2023 Xenium breast data as the '10x Xenium FFPE human breast preview (Janesick 2023)' row (2D)
Same-section DAPI-to-H&E is near-rigid; Xenium-to-Visium is a serial-section problem Used in 2 of the 2D and microscopy registration papers read (STalign 2023; Spotscape 2025). Licence checked on 2026-10-10 at https://github.com/10XGenomics/janesick_nature_comms_2023_companion; derived test pairs shareable: yes; commercial use: yes. data come from the 10x preview dataset page (CC BY 4.0 line on 10x dataset pages, page itself rate-limited) and GEO GSE243280; the repo LICENSE limits the code to use with 10x products, not the dataSame section: DAPI nuclei vs hematoxylin nuclei give dense truth; two people place 20 landmarks on the serial pair
Xenium with H&E and IHC annotation integration (STAR Protocols 2025)Xenium output plus H&E and IHC images to integrate with Xenium at single-cell level, one tissue sample1 sample: Xenium DAPI vs H&E vs IHCAlignment in the provided QuPath project (manual/algorithmic)one imagedirectCC BY 4.0openleft out
one sample (Xenium DAPI, H&E, IHC): a tutorial set, not a test set
Small tutorial-scale set; section relation of IHC (same or serial) unverifiedTwo people place 20 landmarks on each pair
XRF-to-optical microscopy field-of-view localisation setsX-ray fluorescence element maps and optical microscopy of the same or adjacent tissue sectionstwo collections (same-section and adjacent-section); counts unverifiedreference placements (method unverified); companion paper uses scan metadatahiddenunverifiedunknownleft out
its images or its truth are not released
Data release not confirmed.scan stage coordinates give the tile-group placement; two people mark 6 landmarks per map
Z-Brain Browser (Marquart 2017)larval zebrafish brains, inter-specimenhundreds of scans10 to 41 landmarks by 3 blinded scorersboth imagesdirectCC0 1.0opentruth to make
732 unregistered 3D larval brain stacks (CC0) with the authors' registered versions; the blinded scorers' landmarks are not in the release (one landmark volume on one reference brain)
Licence checked on 2026-10-10 at https://doi.org/10.5524/100322; derived test pairs shareable: yes; commercial use: yes. row DOI 10.5524/100314 is a wrong record (rice mutants); the paper's supporting data is GigaDB 100322, CC0 per DataCite; this is the ZBB/Marquart set, not Z-Brain proper (Randlett 2015)
ZephIR C. elegans whole-brain trackingVolumetric calcium imaging of freely moving worm, 4 Hz1 long recordingTracked neuronal nucleiboth imagesdirect: Zenodo, one 14.9 GB zip whose members read by HTTP ranges (the worm is 0.14 GB of it)CC BY 4.0 on the Zenodo recordopenin the corpuspriority 2
The zip holds the ZephIR worm (ZM9624: 1,060 volumes of 2x23x512x512 at 4 Hz, 138 MB) whose annotations.h5 marks people's points (provenance JY, SG, JY0, ANTT) on about ten full reference frames (0, 200, 444, 463, 498, 600, 672, 700, 800, 900, at least 161 of the 178 neurons each) and on 10 neurons over all frames, the other 168 tracked by ZephIR (NEIR). It also holds one Nguyen 2017 worm (nguyen2017_AL_w1, 1,536 frames, 79 neurons tracked by the Leifer lab) that registration 1192 lists behind an IEEE DataPort account, plus an unrelated mouse video. Pairs among the reference frames give dense people's landmarks with large deformation, but from one animal (a second for development), so it ranks apart.
ZeroReg3D mouse kidney consecutive sections (normal and db/db)Consecutive 2 um mouse kidney sections, healthy and db/db diabetic; plus human kidney needle biopsies (H&E, PAS, Jones)29 + 39 consecutive mouse sections; 20 human cases x 13 WSIsLandmarks on mouse sections (maker not stated); none stated for human biopsiesboth imageshiddenNot statedunknownleft out
its images or its truth are not released
Public release not stated; code only on GitHub (hrlblab/ZeroReg3D)Two people place 20 landmarks (glomeruli, vessel branches) on 10 adjacent pairs

Large volume · 62

DatasetProblemPairsChecked againstAccessLicenceStatusNotesTo make the truth
7T ex vivo human brain at 100 µm (Edlow 2019)Ex vivo whole human brain, 7T FLASH at 100 µm, four flip angles (15, 20, 25, 30 deg) and synthesized FLASH; plus MNI-normalised version1 brain: flip-angle volumes to each other; native to MNI 2009bExpert-guided MNI normalisation (ANTs then Lead-DBS, manual fiducials, 132 iterations); manual STN, GPi/GPe, red nucleus in native spaceone imagedirectCC0 (Dryad API; OpenNeuro ds002179 CC0)openleft out
one brain: the flip-angle volumes of one specimen and its MNI warp (provider, warps on request) cannot give ten independent pairs
Other flip angles only on OpenNeuro as source data; 200 and 500 µm versions included.Two people place AFID-style 32 landmarks on the 100 µm volume and on MNI; known smooth warp applied chunk by chunk (after recompressing to OME-Zarr); flip-angle pairs are near-aligned by construction, so a known warp gives exact multi-contrast truth
AIND SmartSPIM whole mouse brains (aind-open-data)Cleared mouse brain light-sheet (SmartSPIM, 3 channels) to Allen CCFv3; also raw tiles to stitched volume; between subjects~4,600 raw SmartSPIM sessions and ~5,300 stitched assets in the S3 listing (2026-10-10); one subject can have several sessionsNo manual truth in the bucket. Each stitched asset has the pipeline's ANTs warps (ls_ccf / ccf_ls displacement fields + affine) and TeraStitcher XMLs (stage import + computed tile displacements)one imagedirectCC BY 4.0openin the corpus
hundreds of brains, CC BY 4.0, anonymous S3 OME-Zarr (level 3 = 1.1 GB a channel, level 0 = 565 GB); only pipeline CCF warps (provider): two people place ~20 landmarks a brain on 10 brains and on the CCFv3 template at level 3 for a main board; a known warp between two channels of one session (synthetic, apart) waits for the block-wise warp (fidopt refuses warp on a large dataset); the stores carry a t axis prepare.py must drop
Raw unstitched tiles kept beside each stitched asset, so stage coordinates vs stitched result can be compared. CCF warp computed at multiscale 3 then 25 um; expert labels exist for 4 brains in a separate Kitware bucket (own entry).Tiles: TeraStitcher xml_import (stage) vs xml_displcomp (refined) gives a stitching reference; apply a known smooth warp chunk by chunk to the fused OME-Zarr for exact full-res truth; for CCF, two people place ~20 landmarks at level 3-4 and carry them to level 0 by the scale factors.
AIND-Kitware SmartSPIM expert annotationsSmartSPIM mouse brain (sample space) to Allen CCFv3; between subjects/atlas4 brains (631680, 640390, 652504, 652506)Expert binary labels of small structures drawn in SmartSPIM space (facial nerve VIIn, medial habenula, fasciculus retroflexus, anterior commissure olfactory part; thalamus for 652504), 'Annotator1/2' folders; compared with CCF labels by signed surface distanceboth imagesdirectnone stated for the annotations; source SmartSPIM images from aind-open-data are CC BY 4.0asktruth to make
4 brains with expert labels of 4-5 small structures by two annotators, against the CCF annotation: too few pairs and labels in a corner of the brain; fold them into the SmartSPIM landmark plan as a check; the annotations carry no licence (ask)
Bucket s3://aind-kitware-collab; annotations made at reduced resolution. Best small truth found for LSFM-to-CCF on public TB-scale data. Licence checked on 2026-10-10 at http://aind-kitware-collab.s3-website-us-west-2.amazonaws.com/; derived test pairs shareable: unclear; commercial use: unclear. collab page names no licence; the AWS Open Data registry lists aind-open-data as CC-BY-4.0, which covers the images but not clearly the Kitware annotationsAdd 20 landmarks per brain by two raters at level 3, scale to level 0; extend labels to more brains with the same protocol.
Allen human brain MRI + multi-stain histology (OpenNeuro ds003590)Ex vivo MRI vs multiple histological stains, 1 brain1 subjectAuthors' joint registration (not independent)directCC0openleft out
one brain; the truth is the authors' own joint registration
Licence checked on 2026-10-10 at https://openneuro.org/datasets/ds003590; derived test pairs shareable: yes; commercial use: yes. dataset_description.json License CC0 (v1.0.2); the underlying images come from the Allen human brain atlas, whose own terms are non-commercial, so the CC0 grant may not clear upstream rightsTwo people place 10 landmarks on 10 section/MRI slice pairs; the authors' joint registration is a reference only. One brain is a demo, not a test set.
Allen Mouse Brain Connectivity Atlas (serial two-photon) with section-to-CCF alignmentsSerial two-photon tomography of whole mouse brains (140 coronal sections at 100 µm, 0.35 µm pixels) vs Allen CCFv3; section-to-volume and between subjects~2,400 projection experiments (API count, product 5)Allen pipeline: per-section 2D affine (alignment2d), per-brain 3D affine (alignment3d) plus deformable grid registration to CCF; QC'd, automaticone imagedirectAllen Institute Terms of Use (non-commercial by default)non-commercialtruth to make
~2,400 brains through the free API (no account) but Allen terms are non-commercial and the CCF alignment is the pipeline's (provider); people landmarks on 10 brains at downsample 4 would make it main; SmartSPIM (CC BY) does the same job openly
Two people place 20 landmarks on 10 brains at downsample 4 and in CCF; carry to full resolution through the known downsampling; known in-plane warp per section gives exact 2D truth
BANC brain and nerve cord (Bates 2025)serial-section TEM of a whole female fly CNS vs JRC2018F/VNC templates and vs other connectomes (FAFB, FANC, hemibrain, MANC); between subjects1 EM volume vs 2 templates and 4 other EM datasetsJRC2018F template already warped into BANC and other datasets' meshes moved in by elastix + thin-plate spline (BossDB channels *_elastix_tpsreg_240721); method outputs, not independentdirectCC BY 4.0 (Virtual Fly Brain, BANC v626)openleft out
one EM specimen; the bridges to templates and other connectomes are method outputs
EM is public on BossDB; segmentation and proofreading need permission. navis-flybrains BANC-maleCNS 'landmarks' are grid points pushed through template chains.Cell types matched between BANC and FlyWire/MANC (from the BANC annotations) give thousands of soma landmarks on both specimens; verify a sample by eye
Benchtop mesoSPIM datasets (BioImage Archive S-BIAD963)Cleared mouse brain/spinal cord and other samples, mesoSPIM light-sheet; stitching and multiview fusionStitched/fused datasets only (114 files); one unstitched 2x2-tile, 2-channel, 2-arm mouse brain (~50 GB OME-Zarr) shared via GlobusNone (beads and Ronchi slide samples for resolution only)one imagedirectCC BY 4.0openleft out
stitched or fused volumes without any truth or second image; the one unstitched brain is behind a Globus login
Original mesoSPIM data on IDR idr0066. Globus share needs a Globus login. Fails truth.Stage coordinates from the BigStitcher XML as initial truth; overlap matches checked by a person; known smooth warp applied per tile.
BigBrainhistology volume of one brain at 20 µm, and histology to MRI1 brain30 histology to MRI landmarks (Xiao 2019, CC BY 4.0) at 0.3 mmboth imagesdirectresearch only, no commercial use without written consentnon-commercialleft out
planned, but one brain: at most three pairs to MRI templates (PD25, ICBM sym/asym); its truth is expert (AFIDs by 8 raters, 22 manual subcortical labels at 0.3-0.5 mm, Xiao 2019, CC BY 4.0) and set at 0.3 mm, so the 20 um volume adds nothing; an unranked histology-to-MRI demo at most
Chosen for the large-volume track; the landmarks check only at low resolution.
BigBrain 1 µm section scans aligned to the 20 µm volumeCell-body stained sections of BigBrain scanned at 1 µm in plane vs the 20 µm 3D reconstruction; section-to-volume, same brainSelected 1 µm sections (cortical patches, e.g. 120 patches in Julich work)Julich alignment of 1 µm sections into the 20 µm space (method output)one imagedirectBigBrain terms: research only, no commercial use without consent (as for BigBrain)non-commercialleft out
one brain; the 1 um to 20 um alignment is Julich's (provider)
BigBrain itself is already listed; FTP readme says raw 1 µm sections are few and unaligned (rescanned).Zoom-to-overview: downsample a 1 µm section to 20 µm and register to the matching BigBrain slice; truth from the known downsampling plus a known in-plane warp; cell-dense landmarks (Betz cells, layer boundaries) by two people
BigBrain2 (second Julich brain)7,676 Merker-stained 20 µm coronal sections of a second human brain, scanned at 10 µm and at 1 µm; 3D reconstruction; same brainSection-to-section and section-to-volume126 structures annotated on sections (Julich Brain Atlas); reconstruction at 100 µm onlyone imagehiddenunverifiedunknownleft out
its images or its truth are not released
Fails access: full 20 µm reconstruction not released as of the 2025 sources.When released: section stack to the authors' volume as truth of the method, landmarks between BigBrain and BigBrain2 for inter-subject
Bosch 2022 correlative 2P, X-ray and SBEM (Crick warpAnnotations)same mouse olfactory bulb or CA1 tissue: in vivo and ex vivo 2P, lab micro-CT, synchrotron X-ray (Diamond, TOMCAT, ESRF ID19, ID16A XNH) and SBEM; same subject, multimodal and multiscale~40 landmark files over ~8 specimens (C525a/b, C432, C556, Y489, Y391, ...)manual BigWarp landmark pairs per modality pair, some placed by two people (CB, IW); independent blood vessel tracings in 2P, SXRT and SBEM to measure warping accuracy (Zhang 2022)both imagesdirectnone statedaskleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): its licence is not shown to allow commercial use (none stated). It was proposed as: manual BigWarp landmark pairs on both images, 9 specimens: lab micro-CT to synchrotron CT 9-25 pairs (median 15), in vivo 2P to synchrotron 36-148 pairs on 4 specimens; public webKnossos datasets with mags 1-32, anonymous API; licence not stated (ask); below ten specimens, so apart
One of the few public multimodal LM-X-ray-EM sets with independent validation annotations; protocolBLAST repo adds artefact examples. Licence checked on 2026-10-10 at https://pmc.ncbi.nlm.nih.gov/articles/PMC9132960/; derived test pairs shareable: unclear; commercial use: unclear. paper: datasets and annotations accessible through the warpAnnotations repo, which has no LICENSE; its Zenodo snapshot (10.5281/zenodo.6342308) is open access without a licence
BRIDGE whole-brain multi-round microscopy benchmarkMouse whole-brain microscopy: cross-subject and to CCFv3; multi-round FISHRepresentative subsetsRegional annotations; nuclei validation setdirectCC BY 4.0openneeds a request
only representative subsets are deposited (0.37 GB: two brains with atlas-propagated labels, 50 round-to-round sub-volumes with nuclear landmarks whose making is not described); the TB-scale brains are not
Full TB-scale data not depositedRounds share a DAPI channel: nuclei matched across rounds by a program and verified by a person give dense point truth; regional annotations on both as a weaker check for cross-subject pairs.
CANVAS whole-brain light-sheet benchmark (2026)SmartSPIM whole mouse brains immunolabelled for NeuN, cFos, PV, TH, IBA1, GFAP; single brains, cell detectionNone for registrationCell centroids in training ROIs (45,745); test centroids hiddenone imagedirectCANVAS Dataset licence agreement (LifeCanvas): non-commercial, no redistributionnon-commercialleft out
single brains for cell detection, no registration pair or truth; the agreement signed in the site's app is non-commercial with no redistribution; SmartSPIM covers brains to CCF openly
Fails the pairs criterion; candidate source of big brains for CCF registration tests Licence checked on 2026-10-10 at https://canvas.lightsheetdata.com; derived test pairs shareable: no; commercial use: no. agreement (in the site's app): free for academic and non-commercial research; 'No Redistribution of the Dataset'; derivatives non-commercial only; commercial licensing via data@lifecanvastech.comRegister to Allen CCF and score region labels made by a person on 10 slices per brain; or 20 landmarks by two people
Chung lab SHIELD human brain slabs (DANDI 000108)Light-sheet of cleared 1-4 mm human brain slabs (whole hemisphere and blocks): chunk stitching, slab-to-slab (UNSLICE), multi-stain same slab; same subject5 subjects; thousands of chunk pairs; consecutive slabs; NeuN/YOYO/lectin stains of the same slabPer chunk ChunkTransformMatrix (stage offset, voxel size) in the sidecar JSON; no inter-slab truth deposited (paper used fiber endpoints, 1.1-1.3 um error on small ground-truth sets not on DANDI)one imagedirectCC BY 4.0opentruth to make
5 subjects only, ~100 GB a chunk at level 0; stains of one chunk aligned by construction allow a known warp (synthetic, waits for the block-wise warp); slab-to-slab vessel endpoints matched by a program and checked by people would be semi-automatic
BIDS+OME-Zarr; Neuroglancer dashboard stitches chunks with the stage rigid transform only. Stains in one session share the stage transform, so they are aligned by construction.Stitching: stage offsets as initial truth, refine with overlap matches and check by bead-free vessel continuity; multimodal: apply a known smooth warp to one stain and register it back to another stain of the same chunk (exact truth); slabs: vessel endpoints matched on cut faces by program, verified by a person.
CloudReg cross-modal brain volumes (Chandrashekhar 2021)Cleared mouse and rat brains (CLARITY, iDISCO, SHIELD light-sheet), mouse micro-CT, macaque and human MRI to atlasesNot statedLandmarks placed by one person (J. Rosenblum) for accuracy; counts not statedboth imagesrequestno data licence; code Apache-2.0askneeds a request
the data needs a request or a data use agreement
Data available from the corresponding author on request only Licence checked on 2026-10-10 at https://cloudreg.neurodata.io/license.html; derived test pairs shareable: unclear; commercial use: unclear. licence page covers the CloudReg software only; paper data available from the corresponding author on requestAsk for the landmark files; otherwise two people place 20 CCF landmarks on each released brain
CynoBrain raw multimodal macaque datacynomolgus macaque in vivo 3T MRI, ex vivo 9.4T MRI and whole-brain autofluorescence imagingsubject count unverifiednone stated (atlas inputs)one imagedirectCC BY 4.0opentruth to make
in vivo MRI, ex vivo MRI and whole-brain optical of macaques, CC BY 4.0, 246 GB; the number of animals and whether ScienceDB needs a login could not be read (JavaScript page): check first; landmarks by two people across modalities
In vivo to ex vivo to optical chain.vessel and sulcal landmarks by two people across modalities
Deisseroth CLARITY COLM brains (NeuroData tomer15)CLARITY light-sheet (COLM) mouse brains to Allen atlas and to each other; between subjects10 brains on S3 (Cocaine, Control, Fear groups); paper used 4 complete onesKutten 2016 placed 55 landmarks (28 surface, 27 internal) on ARA and CLARITY brains in DiffeoMap; landmark files not found in ndreg repo or bucketdirectODC-By 1.0opentruth to make
10 brains on S3 (several incomplete), ODC-By, precomputed 7 scales; Kutten 2016's 55 landmarks on 4 brains are not public (ask JHU) and four brains are too few: place landmarks on 10 brains + ARA
Old pipeline, behaviorally challenged brains, some brains miss data. Landmarks would need to be asked from the authors.Ask JHU for the 55 landmarks; otherwise two raters place 20-30 landmarks at scale 4-5 and scale to full res; known smooth warp applied per chunk on the precomputed volume.
DevCCF paired MRI and LSFM specimens (Brain Image Library)Ex vivo in-skull MRI (DWI/T2/FA/ADC, 31.5-50 um) to cleared whole-brain LSFM (1.8x1.8x5 um, 3 channels) of the same mouse; P14 and P56 (and other ages unpaired)>=9 specimens with both MRI and LSFM found by BIL metadata (P14 x5, P56 x4); 66 MRI and 70 LSFM samples in the paperNone per specimen. Templates were joined by landmark-assisted ANTs registration (misaligned regions segmented in Avizo); landmarks are region masks, not points, and not per specimenone imagedirectCC BY-SA 4.0 for the paired sets (other Kim lab LSFM CC BY-NC-SA 4.0)opentruth to make
at least 9 mice with in-skull MRI and cleared LSFM of the same brain (large real deformation), CC BY-SA 4.0, BIL free without account (its download server timed out from here); two raters place 20-30 landmarks at 10-25 um; the 3 LSFM channels also allow a known warp
Same-specimen MRI-LSFM is rare at this scale; clearing shrinkage and brain extraction give a large real deformation.Two raters place 20-30 landmarks (vessels, ventricles, tracts) on MRI and on LSFM at 10-25 um, carried to full res by scale; DevCCF labels warped onto both as a weaker check.
Duke Mouse Brain Atlas: MRH 15 µm + light-sheet of the same brains (Mansour 2025, Tian 2022)Perfusion-fixed C57BL/6J mouse brains: in-skull MR histology (DTI 15 µm, ~252 GB 4D) then cleared light-sheet (1.8x1.8x4 µm, ~300 GB per channel) of the same brain; micro-CT skull; same specimen, multimodal; also to CCFv35 atlas specimens (MRH-LSM); more in Tian 2022 experiments 1-3Fiducial landmarks on MRH and LSM (n=50, 175, 200 sets in Tian 2022, vessel bifurcations etc.), 15-20 manual init pairs per specimen; RCCFv3 labels by two neuroanatomistsboth imagesrequestCC BY-NC-SA 4.0non-commercialneeds a request
the data needs a request or a data use agreement
Login on request (CIVMImageSpace). Whether the landmark files are released is unverified. Related: HiDiver entry (Johnson 2023, same portal).If the 50-200 landmark sets are released, use them as held-out truth; otherwise two people place 30 vessel bifurcations on MRH and LSM at 4x downsample, carried to full resolution
EASI-FISH multi-round expansion FISH example data (Janelia)Expanded mouse lateral hypothalamus slab imaged by light-sheet in rounds 3, 4 and 5 (DAPI in every round); same tissue across rounds, tiled with 8% overlapRounds 3-4-5 (round 3 to 5 suggested for registration testing)None; bigstream deformable registration in the authors' pipelineone imagedirectCC BY 4.0openleft out
one tissue slab imaged in rounds 3-5: one specimen
Gel deformation between rounds is real and non-rigid.DAPI nuclei segmented in each round (Cellpose) and matched to give point truth, then checked by a person on 200 nuclei; tile overlaps give stitching truth; known warp applied chunk by chunk to round 3
EMIT WSI and MCMICRO exemplars (CyCIF multi-cycle)Cyclic immunofluorescence of tonsil whole slide and TMAs: many cycles, each with a DNA channel; same section across cycles; raw tiles stitched with ASHLARCycle-to-cycle (exemplar-002: 10 cycles; exemplar-001: 3 cycles); tile-to-tileNone; ASHLAR uses tile overlaps and DNA channel; stage positions in image metadata (unverified per set)one imageaccountSite content CC BY-NC 4.0; EMIT on Synapse (terms unverified)non-commercialneeds an account
the download needs an account
Exemplars on public S3 (direct); EMIT via Synapse (account).Stage coordinates plus overlaps give stitching truth; DNA nuclei matched across cycles give point truth after a person checks 100; known warp to one cycle gives exact truth
Ex vivo corneal stromal nerve plexus, hexapod confocal mosaics (Zenodo 17544799)confocal volumes of corneas positioned by a six-axis hexapod, stitched into mosaics; porcine, murine, ovine, humanseveral corneas, many tiles eachhexapod stage positions (raw tiles with coordinates unverified); manual nerve annotationsdirectCC BY 4.0openleft out
one fused volume per cornea, no raw tiles, no pair; nerve masks by a network (manual only on one crop)
stitched mosaics are given; check for raw tilesstage coordinates as the stitching truth; nerve branch points across overlaps checked by a person
ExaSPIM (Allen AIND)expansion light-sheet, whole mouse brain, sub-µmmany brainsno landmarksdirectCC BY 4.0opentruth to make
planned, but no correspondence made by people and a single channel (no pair aligned by construction for a known warp); level 0 ~67 TB a brain (21605x28886x54049 uint16), pyramid to level 7; CCF warps are the pipeline's; landmarks by people on 10 brains at level 5-6 would make it main, SmartSPIM first
Chosen for the large-volume track; truth would be synthetic.Tile stage coordinates check stitching; a known smooth warp on a whole brain gives synthetic truth at scale; two people place 20 landmarks on 5 inter-brain pairs at 25 um, carried to full resolution by scale.
Extremely Brilliant Brain (EBB, HiP-CT whole human brain + 3T MRI)HiP-CT (BM18) of one whole post-mortem human brain at 7.72 µm vs 3T T2 MRI of the same brain; zooms (VOIs) to 1.3 µm; same subject, multimodal and multiscale1 brain: HiP-CT to MRI, HiP-CT to BigBrain, VOIs to overviewAuthors' rigid MRI-to-HiP-CT registration and voluba affine to BigBrain (4x4 given); FastSurfer labels on binned HiP-CT, HippUnfold on MRI, left hippocampus refined at full resolutionone imagedirectCC BY 4.0openleft out
one brain; only the authors' rigid and affine registrations
Donor LADAF-2021-17; 2026 release on top of the Human Organ Atlas (already listed). No zoom-to-overview transforms published for this brain, no landmarks.Place 20-30 landmarks on the 61.8 µm level (29 GB) and on MRI, carry to full resolution; label hippocampus on both; known smooth warp applied chunk by chunk to the 13.7 TiB volume; VOIs to overview by rigid search with stage positions as start
FAFB alignments v14, v14.1 (FlyWire) and v15 (same sections)serial-section TEM section-to-section alignment, one fly brain; same raw sections aligned three times (AlignTK/EM_aligner v14, SEAMLeSS v14.1, SEAMLeSS + vector voting v15)~7000 consecutive section pairs; 3 whole-volume alignments of the same datano ground truth; chunked Pearson correlation between neighbours plus 300 low-correlation sites per version classified by a human expert; deformation fields v14 to v14.1 from the Seung lab (fafbseg)directv15: CC BY-NC 4.0; v14 and v14.1 licence not stated (unverified)non-commercialleft out
one specimen aligned three times, no truth
Rare chance to compare whole-brain alignments of identical raw data; unaligned raw tiles not found public. temca2data.org also serves the 2018 paper alignment. Used in 1 of the 2D and microscopy registration papers read (Popovych 2024 (SEAMLeSS)).Take v15 sections, apply known smooth per-section warps plus section jumps chunk by chunk at full resolution, score recovery; or two people mark the same mitochondria/synapses across 5-section spans at 300 sites in each version
FAFB whole fly brain EM to JFRC2013 template (ELM landmarks)ssTEM of one adult female fly brain vs light-microscopy template brain (JFRC2013); between subjects, EM to LM1 pair (FAFB v14 vs JFRC2013); plus FAFB to JRC2018F bridge135 manual landmark pairs (lm-em-landmarks_v14.csv) placed in BigWarp by the Bock and Saalfeld groups; FAFB to JRC2018F transform made by ANTs on blurred synapse predictions (Bogovic 2020)both imagesdirectFAFB images CC BY 4.0 (FlyWire release on BossDB); ELM landmarks GPL-2.0openleft out
one pair (135 manual landmark pairs, expert): a fine single case, not a dataset
Landmarks were placed at template resolution; the EM side is in nm. Earlier sets for FAFB12/13 also kept. Truth is the fitting set, so hold some out. Licence checked on 2026-10-10 at https://doi.org/10.60533/BOSS-2020-F81C; derived test pairs shareable: yes; commercial use: yes. the FlyWire BossDB record (FAFB v14.1 realigned images) is CC BY 4.0 in DataCite; the original temca2data.org v14 release was not checked; landmark repo saalfeldlab/elm is GPL-2.0Hold out a third of the 135 pairs; add soma positions of identified neurons matched between FlyWire and LM lines (NBLAST) as extra pairs; a second person re-places 20 landmarks to get inter-rater error
FANC female adult nerve cord (Phelps 2021) to JRC2018 VNC female templateGridTape serial-section TEM of a female fly VNC vs JRCVNC2018F light template; between subjects, EM to LM1 pair to templateelastix transform FANC to JRCVNC2018F by Jasper Phelps (navis-flybrains); how it was made not seendirectCC BY 4.0 (FANC EM on BossDB); JRC2018 VNC template on Janelia figshare (CC BY 4.0 on the records checked)openleft out
one specimen; the template transform is a method output
Proofread segmentation needs FANC community access; EM image is open. Licence checked on 2026-10-10 at https://doi.org/10.60533/BOSS-2021-T21Z; derived test pairs shareable: yes; commercial use: yes. BossDB DOI rights CC BY 4.0; fanc/flybrains code GPL-3.0 is separateLeg and wing motor neurons identified in both FANC and LM driver lines (Azevedo 2024 atlas) give nerve-root and soma landmarks; two people place 20 neuropil landmarks at template resolution
Fish1 larval zebrafish CLEM connectome (Petkova 2025)live confocal (vglut2a, gad1b, kdrl vessels) vs serial-section EM of the same 7 dpf zebrafish; same subject, LM to EM, whole brain1 whole-brain pair9,421 LM-EM cell correspondences from iterative point-cloud matching after manual coarse landmarks (vessels, unique neurons); 70,000 overlay images reviewed by people; 41,175 matched labelled neuronsboth imagesdirectno named licence; 'Data Usage & Publication Policy' (open access, cite Petkova 2025)askleft out
one fish; the LM-EM matches and raw confocal are not public; no licence
Released confocal looks already warped into EM space; raw confocal and the landmark table not found public (unverified). A clem_labels layer is public. Licence checked on 2026-10-10 at https://fish1-release.storage.googleapis.com/data_policy.html; derived test pairs shareable: unclear; commercial use: unclear. policy invites independent analysis and publication with citation, no co-authorship required, but grants no licence for copying or redistributionAsk for raw confocal and the 9,421 pairs; meanwhile apply a known smooth warp to the warped confocal and use clem_labels matched cells as label truth
FlyWire / FAFB whole fly brain ssTEMserial section TEM of a whole adult Drosophila brain (FAFB, Zheng 2018), re-aligned (FAFBv14.1), section alignment~7,000 sectionsproof-read whole-brain neuron segmentation; CREMI realigned volumes released by FlyWireboth imagesdirectFlyWire data CC BY 4.0 (unverified per release)openleft out
one specimen; section alignment only, no pairwise truth
Original and realigned versions both exist, a rare before/after alignment pairneuron skeleton continuity across sections as a dense check
fMOST mouse brains (Brain Image Library)fMOST whole mouse brains, inter-animal; downsampled sets with mBrainAlignermanylandmarks unverifieddirectCC BY-SA 4.0opentruth to make
many brains, CC BY-SA 4.0, BIL (server timed out from here); mBrainAligner releases no manual truth (only ~50 um sample brains and network masks): landmarks by people on 10 brains + CCF
Licence checked on 2026-10-10 at https://www.brainimagelibrary.org/download.html; derived test pairs shareable: yes; commercial use: yes. BIL 'Distribution Licence: BIL Data is distributed under the Creative Commons Attribution-ShareAlike 4.0 International License'; share-alike; download directory did not answer from hereTwo raters place 20-30 landmarks on 10 brains and the CCF at 25 um (as in mBrainAligner); first check whether mBrainAligner's landmarks are released.
H01 human cortex (Shapson-Coe 2024)petascale serial section EM of human temporal cortex, section alignment and stitching~5,000 sections, 1.4 PBaligned volume, proof-read neurons and synapses (subset)both imagesdirectCC BY 4.0 (per release page, unverified)openleft out
one specimen, aligned volume; only synthetic warps on crops possible
Raw tiles before alignment availability unverifiedknown warps on aligned crops
Hemibrain FIB-SEM (Scheffer 2020) and hemibrain to JRC2018F bridgeisotropic FIB-SEM of half a female fly brain vs JRC2018F light template; between subjects, EM to LM1 pair (hemibrain vs JRC2018F)transform JRC2018F_JRCFIB2018F.h5: ANTs on a synthetic synapse-density image, then manual BigWarp correction (Bogovic 2020); landmarks not publisheddirectdata CC BY 4.0; transform CC BY 4.0 (figshare 11871699)openleft out
one specimen; the bridge is ANTs plus manual correction, landmarks not published
Isotropic 8 nm: the best large source for exact synthetic serial-section truth.Cut the isotropic volume virtually into 40 nm sections and apply known per-section warps chunk by chunk, so ssEM alignment gets exact 3D truth; for EM-LM, soma positions of cell types matched between hemibrain and FlyWire (Schlegel 2024) give many landmark pairs
HiDiver merged MRH and light-sheet (Johnson 2023, Duke CIVM)15 um in-skull MR histology (GRE, DTI) to cleared LSM (1.8 um, NeuN etc.) of the same mouse brain; same subjectExperiments 1-3 of the paper (C57BL/6 and BXD mice; exact count unverified)Authors report 22 +- 14 um average alignment from 11 vessel measurements; per-specimen landmark files not seenaccountCC BY-NC-SA (as stated in the paper)non-commercialneeds an account
the download needs an account
Access page is a login portal; whether a guest login works is unverified.Vessel bifurcations found by program in both MRH and LSM, verified by a person (20 per brain); carry from 15-25 um to LSM full res.
Hildebrand 2017 whole larval zebrafish ssEM with 2P and Z-Brainserial-section SEM of a 5.5 dpf larval zebrafish vs in vivo 2P (GCaMP5G) of the same fish and vs Z-Brain / Zebrafish Brain Browser atlases1 same-fish 2P-EM pair; 2 atlas-EM pairslandmark thin-plate splines in BigWarp by the authors (nuclei patterns for 2P; region boundaries, ganglia, midline for Z-Brain); landmark files not seenboth imagesdirectOpen Database License + Database Contents Licenseopenleft out
one fish; the authors' landmark files are not public
Warped Z-Brain stacks are public in EM space; same-fish 2P stack availability not confirmed.Re-place 20-30 landmarks on Z-Brain vs EM by two people (Mauthner cells, nucMLF, ganglia); or warp the public EM-space Z-Brain with a known field
Human brain cell census BA 44/45 (DANDI 000026)MRI, light-sheet, PS-OCT, histology of same human brain blocksUnverifiedNone (PS-OCT used to register)one imagedirectCC BY 4.0opentruth to make
12 subjects with ex vivo 7T MRI (~4.9 GB a volume) and PS-OCT (30x30x35 um, 2-4.5 GB) of the same Broca block, CC BY 4.0, DANDI anonymous; two people place 10-20 landmarks (vessels, cortical points) a pair: ten independent pairs, main board; SPIM stains (~28 GB OME-Zarr each) also allow a known warp
Two people place 10 landmarks (vessels, cortical points) on PS-OCT, light-sheet and MRI of 10 blocks (inter-observer error reported).
Human hippocampus dMRI + 3D-PLI + TPFM (EBRAINS, Axer/Poupon/Costantini 2020)One post-mortem hippocampus: 11.7T T1/T2 at 200 µm and dMRI at 300 µm, blockface, 530 3D-PLI sections at 1.3 µm, TPFM at 0.44x0.44x1 µm on selected sections; same specimen, multiscaleMRI to blockface to 3D-PLI to TPFM (zoom chain)Authors' co-alignment via blockface; no landmarks statedone imagedirectCC BY-NC-SA 4.0 (DataCite)non-commercialleft out
one specimen; the co-alignment is the authors'
Blockface is aligned by construction to each section: known warp of a PLI section gives exact truth to blockface; vessel and fibre landmarks between PLI and TPFM by two people
Human Organ Atlas (HiP-CT)phase-contrast CT, whole organs at ~20 µm with zooms down to 0.65 µm, same organ296 images, 23 donorseach zoom registered to its overview, rigid transforms in the metadataboth imagesdirectCC BY 4.0opentruth to make
planned, but a zoom and its overview differ by one similarity transform (the provider's MI registration in the metadata): nothing left for a deformable method after alignment; inter-donor organ pairs at overview level 2-3, with labels or landmarks drawn by people, would make it main; public GCS bucket, levels 0-4
Chosen for the large-volume track; OME-Zarr in a public bucket.
Janelia MouseLight imagery and registrationWhole mouse brain serial two-photon tile mosaics to Allen CCF; neuron tracings in native space; between subjects68 imaged brains (images/), 45 with CCF registration foldersPer brain: CCF displacement field (Transform.<date>.h5, ~65 GB), 10 um downsampled volume and its registered version, CCF meshes back-warped to sample space; registration was BRAINSFit affine + iterative TPS landmark registration in 3D Slicer (landmark files not seen); tile stitching control points in <date>.ymlone imagedirectCC BY 4.0opentruth to make
45 brains with a 10 um NRRD (0.65 GB) and the CCF transform (provider), CC BY 4.0, anonymous S3; the TPS landmarks are not published: two people place 20 landmarks on 10 NRRDs + CCF; no chunked reading needed at 10 um
Bucket janelia-mouselight-imagery; raw tiles referenced in the yml, not confirmed in the bucket. Neuron tracings (consensus of 2+ annotators) exist in native and CCF space.Recover the TPS landmarks if shared (unverified); else two raters place 20 landmarks on the 10 um NRRD and the CCF template; traced neurons give curve-to-curve checks after warping; known smooth warp applied block by block on the octree.
Kasthuri 2015 mouse neocortex ssEMserial section SEM volume of mouse neocortex, aligned; section alignment~2,000 sections (aligned volume)partial dense neuron and synapse annotationsboth imagesdirectODC-By v1.0 (NeuroData); CC BY 4.0 (BossDB DOI)openleft out
one specimen, aligned volume; only synthetic warps on crops possible
Raw unaligned tiles availability unverified Licence checked on 2026-10-10 at https://neurodata.io/data/kasthuri15/; derived test pairs shareable: yes; commercial use: yes. NeuroData page: 'Data on this site are licensed under ODC-By v1.0'; BossDB record 10.60533/BOSS-2015-IOQM is CC BY 4.0 in DataCiteknown warps on aligned crops; label continuity across sections
Kuan 2020 X-ray holographic nano-tomography (XNH) with correlative EMsynchrotron XNH of fly brain, VNC, leg and mouse cortex; correlative serial TEM of the same mouse PPC block and fly leg nerve sections; same subject, X-ray to EMmouse cortex XNH-EM pair; fly leg nerve XNH-TEM sections; 10-12 overlapping leg XNH scans stitchedXNH and EM aligned by the authors in BigWarp (landmarks not seen public); EM segmentation used as truth for 2D XNH segmentationdirectCC BY 4.0openleft out
one specimen per modality pair; the BigWarp landmarks are not public
Correlative EM of mouse cortex is served through the Lee lab CATMAID, not BossDB; download not confirmed. Licence checked on 2026-10-10 at https://doi.org/10.60533/BOSS-2020-LCIP; derived test pairs shareable: yes; commercial use: yes. BossDB project DOI, DataCite rights CC BY 4.0 (bossdb.org page is JavaScript-only)Two people place 20-30 landmarks (somata, vessels) on XNH and EM; leg scan overlaps give stitching truth from repeated content
LINC human/macaque (DANDI 001278/001372)dMRI, PS-OCT, light-sheet, histology of same brainsUnverifiedNone statedone imagedirectCC BY 4.0openleft out
3 human subjects (DANDI summary) and a 67 GB macaque set: fewer than ten subjects, no truth
Two people place 10 landmarks (vessels, cortical points) on PS-OCT, light-sheet and dMRI of 10 blocks (inter-observer error reported).
Male CNS connectome (Janelia/Google 2025, maleCNS v1.0)FIB-SEM of a whole male fly CNS (brain, optic lobes, VNC) vs JRC2018M template; between subjects; also vs FAFB/FlyWire and BANC1 pair to template; bridges to 3 EM datasetstransform MaleCNS_JRC2018M.h5 (16 GB, figshare 24002001); navis-flybrains maleCNS-FAFB/FlyWire/BANC point sets are grid samples through other transforms (derived), plus a few manual neck pointsdirectCC BY 4.0openleft out
one specimen; template and connectome bridges are method outputs
v0.9 Oct 2025, v1.0 June 2026. Neuropil ROIs were initialised from JRC2018M ROIs.Soma positions of the ~11,700 cell types matched to FlyWire (sexual dimorphism paper) as cross-specimen landmarks; isotropic volume also suits synthetic section-warp truth
MANC male adult nerve cord (Takemura 2024) to JRC2018 VNC male templateFIB-SEM of a male fly ventral nerve cord vs JRCVNC2018M light template; between subjects, EM to LM; also MANC vs FANC (EM to EM, different flies)1 pair to template; 1 pair to FANCbridging transform JRC2018VncM_MANC.h5 (Saalfeld lab, method not stated); MANC-FANC 507 point pairs in navis-flybrains (a regular grid in FANC, so likely derived from another transform)directCC BY 4.0 (data and transform, figshare 14991978)openleft out
one specimen; transforms are method outputs
No independent truth; transforms are method outputs.Soma and nerve-entry points of motor neurons and cell types matched between MANC and FANC (and MANC and templates via split-GAL4 lines) as landmark pairs, checked by a person
Marmoset Brain Architecture Project (Lin 2019) histology with in vivo and ex vivo MRIMarmoset brains cut in ~1,700 20 µm sections (Nissl, myelin, tracer fluorescence) with in vivo and ex vivo MRI of the same animal; section stack to own MRI, then to atlasSeveral brains (count unverified)Authors' MRI-guided 3D reconstruction (Lee 2018 pipeline); no landmarks statedone imagedirectnone stated on the marmoset pageasktruth to make
several marmosets with section stacks and own MRI, but the licence is not stated (ask) and the download path is unverified; landmarks by two people on sections and MRI
Host and download path for raw sections unverified. Licence checked on 2026-10-10 at https://data.brainarchitecture.org/pages/marmoset; derived test pairs shareable: unclear; commercial use: unclear. page describes atlas, raw MRI/CT/histology and downloads with no data licence (earlier GPL v3 text looked like generic site code)Section stack to own MRI is same-subject 2D/3D; two people place 20 landmarks on 5 sections and MRI; known per-section warps give exact stacking truth
MICrONS mm³in vivo two-photon stack to electron microscopy, mouse visual cortex1 volume pairmanual fiducials and 19,181 manual cell matchesboth imagesdirectAllen Institute non-commercial licence for the code; imagery licence unverifiednon-commercialleft out
one pair (manual fiducials and 19,181 manual cell matches, expert)
Used in 1 of the 2D and microscopy registration papers read (Popovych 2024 (SEAMLeSS)).
MICrONS Phase 1 layer 2/3 (pinky100) 2P-EM coregistrationin vivo two-photon structural stack vs serial-section EM of the same mouse V1 layer 2/3 volume; same subject, LM to EM1 volume pairCSV of co-registration points between 2P stack and EM, with a notebook; count and who placed them not seen (unverified)both imagesdirectCC BY 4.0 (MICrONS terms)openleft out
one pair; who placed the points is not stated
Separate from the listed MICrONS mm3. Raw 2P movies (90 GB) on request; large downloads ask to contact MICrONS.Hold out part of the points; add soma centroids matched between the 2P stack and the EM soma table, checked by a person
Multi-round SHIELD sections of human and primate brains (DANDI 000674)4 mm coronal brain sections stained in several rounds and imaged by oblique light sheet, round to round of the same sectionhuman, macaque and marmoset sections (count unverified)none; nuclear stain repeated across roundsdirectCC BY 4.0openleft out
one subject per the DANDI summary (9 files, 12.8 GB); no truth
SHIELD-processed; related to the SHIELD slabs already listed (DANDI 000108)nuclei in the shared channel matched by a program, checked by a person
Multimodal postmortem infant brainstem (DANDI 000874)MRI, OCT and histology of same brainstemsUnverifiedNone statedone imagedirectCC BY 4.0openleft out
one subject (4 samples) per the DANDI summary
Two people place 10 landmarks (vessels, nucleus edges) on MRI, OCT and histology of each brainstem (inter-observer error reported).
Neutron microscopy and X-ray micro-CT of human bone (Zenodo 7743859)3D neutron microscopy vs lab X-ray micro-CT of the same bone samplescount unverifiednonedirectCC BY 4.0openleft out
one bone sample (two 9-18 GB TIFFs), no truth
rigid multimodal problemHaversian canal branch points marked in both volumes by two people
OpenOrganelle CLEM: jrc_cos7-11 and aic_desmosome-2cryo-SIM (and PALM for COS-7) fluorescence vs FIB-SEM of the same frozen cell; same subject, LM to EM2 cells (COS-7 ER/mito SIM+PALM; A431 488/560/640 SIM)LM registered into EM space by John Bogovic; aic_desmosome-2 has raw LM and LM_transformed, so the transform is recoverable; method not seendirectCC BY 4.0 (figshare)openleft out
two cells; the LM-to-EM transform is the provider's
Only these two OpenOrganelle datasets have LM layers in the zarr; other lm/ folders are empty.Two people place 20 landmarks (mitochondria tips, desmosome plaques) on raw LM and EM; or apply a known smooth warp to the registered LM for exact truth
Orion same-section H&E and 18-plex IF, colorectal cancer (Lin 2023)Whole-slide multiplex immunofluorescence (Orion) and H&E on the same tissue section, 74 colorectal resections plus TMA; multimodal, same section74 WSI pairs (H&E to IF) plus TMA coresSame section: aligned by construction up to restaining shift; authors registered H&E to IF with ASHLAR/PALOMboth imagesdirect: anonymous S3 bucket lin-2023-orion-crc (pyramidal OME-TIFF, read by ranges)CC BY 4.0 (HTAN open-access imaging data); MIT on the Zenodo code and index recordopenin the corpuspriority 3
The public bucket s3://lin-2023-orion-crc lists anonymously 41 case folders (CRC01 to CRC40, CRC33 in two), each with the IF image (112.8 GB for CRC01), the H&E already registered to it ('-registered.ome.tif', 1.03 GB), a cell mask and a table. With the H&E delivered in the IF frame, the only truth is a known warp (synthetic, ranked apart): Hoechst against H&E at a pyramid level as a 2D dataset, or the full slides once the block-wise warp exists. HTAN states CC BY 4.0 for its open imaging; the Zenodo record 7637988 is the MIT code and index release.
Code MIT; derived data on Zenodo 7637655. Licence checked on 2026-10-10 at https://humantumoratlas.org/data-access; derived test pairs shareable: yes; commercial use: yes. HTAN data access page: open-access imaging data (IDC, SB-CGC) are CC BY 4.0; same dataset as the ORION-CRC rowExact multimodal truth: take the PALOM-aligned pair, apply a known smooth warp to the H&E chunk by chunk; nuclei centroids (Hoechst vs hematoxylin) as dense check points
RASI serial-section microscopy and synchrotron CT of an LTCC package (BAM)serial-section reflected-light optical stack vs multi-energy HDR synchrotron CT of the same ceramic package1 specimen: 176 sections (12480 x 18582 x 177 voxels) vs a CT volume (Zenodo 17234620)none; silver tracks and vias are natural markers; per-section stage height logdirectCC BY 4.0openleft out
one ceramic specimen; no truth yet
sections aligned in-plane by SIFT; a correlative study is announcedvia centres found by a program in both volumes, checked by a person
SenNet + HOA kidney vasculature (Kaggle 2024)HiP-CT of whole human kidneys at ~50 µm (binned from ~25 µm) plus a high-resolution VOI (kidney_1_voi, ~5.2 µm); same organ at two scales3 training kidneys; 1 VOI-to-overview pair (kidney_1); 2 test kidneysVessel labels by curators (dense on kidney_1 and part of kidney_3, sparse elsewhere); labels on both the overview and the VOIboth imagesaccountCC BY 4.0openneeds an account
the download needs an account
VOI is a rigid subset of the overview; transform not published. Competition data also on a Google Drive, Zenodo deposit planned. Licence checked on 2026-10-10 at https://doi.org/10.5281/zenodo.19685382; derived test pairs shareable: yes; commercial use: yes. Zenodo record by Jain, Walsh et al. holds all competition data (HiP-CT images, labels, whole-kidney test set); no Kaggle account needed; Kaggle rules page itself not readableRigid VOI-to-overview registration checked by vessel-label overlap (labels exist at both scales); vessel bifurcations as landmarks found by a program on the labels, verified by a person
Synchrotron micro-CT time series of dehydrating grapevine petiolessynchrotron micro-CT of grapevine petioles at several dehydration stages, same sampletime series (count unverified)42 manual semantic masks (time points unverified)directCC BY 4.0openleft out
two time series; the 42 manual masks are 2D slices for segmentation, not pairwise truth
ALS beamline 8.3.2vessel-lumen centroids tracked by a program, checked by a person
Synchrotron mouse brain (Rodgers 2025)0.65 µm synchrotron CT, whole mouse brain1noneone imagedirectCC BY 4.0openleft out
one volume, no second image
Licence checked on 2026-10-10 at https://api.datacite.org/dois/10.25493/JWM7-2G1; derived test pairs shareable: yes; commercial use: yes. EBRAINS dataset 'Mosaic microtomography of an entire mouse brain' v2.0, DataCite rightsOne volume, no second image: only tile coordinates for stitching or a known smooth warp (synthetic).
Time-lapse helical CT of fatigue damage in GFRP (Manchester)lab X-ray CT of a glass-fibre composite at several fatigue stages, same specimentime-lapse seriesnonedirectCC BY 4.0openleft out
one specimen at three fatigue stages, no truth
Wang et al., Materials 2018fibre-break centroids found by a program then verified by a person; or a known warp
TubeMap / ClearMap 2 vasculature datasets (OSF)iDISCO+ whole mouse brain hemispheres: raw light-sheet tiles to stitched volume; autofluorescence scan to signal scan; to atlas2 hemispheres (Brain-44R, Brain-39L): tiles, autofluorescence stack, vessel graphNone; stitching by WobblyStitcher and atlas alignment by Elastix in ClearMap, outputs not depositedone imagedirectnone statedaskleft out
two hemispheres, no truth, no licence
Tile file names carry the grid index [row x col]; stage positions in OME metadata (unverified). Licence checked on 2026-10-10 at https://osf.io/sa3x8/; derived test pairs shareable: unclear; commercial use: unclear. OSF project public with empty node_license (row URL uses a view-only link)Stage grid + nominal overlap as initial stitching truth; vessels crossing tile seams found by program and checked by a person; known smooth warp applied per tile for exact stitching truth.
Uninterrupted time-lapse lab CT: mung bean germination and barite precipitationlab CT of a germinating mung bean (54 tomograms) and of barite growing in porous media54-volume series plus a golden-ratio projection seriesnonedirectCC BY 4.0openleft out
one specimen each; growth is not one-to-one
growth is not one-to-one, so only the porous medium suits point truthgrain contact points tracked by a program and checked by a person in the barite series
VesSAP / DISCO vascular brains (Erturk lab)3DISCO-cleared whole mouse brains (vessels) registered to Allen atlas; between subjects (C57BL/6J, CD1, BALB/c)Several whole brains (count unverified)Atlas-registered volumes and segmentations (author pipeline), no manual landmarksone imagedirectCC BY-NC 4.0non-commercialtruth to make
several brains (count unverified), CC BY-NC 4.0, SFTP with public credentials; only the authors' atlas registration: vessel landmarks by people on 10 brains + atlas
Download via SFTP (host/user/password printed in the public guide PDF). vDISCO whole-body and SHANEL human organ data: only videos online, raw data not found.Vessel bifurcations as landmarks found by program on both brains and atlas-aligned volume, verified by a person; known warp on full-res tiles.
Visible Humancryosection, CT, MR of the same male and female cadavers2 bodiesno public landmarksone imagedirectfree since 2019 (NLM), not a CC licenceopenleft out
two bodies: two pairs at most
Two people place 20 bony landmarks on the cryosection volume and on CT and MR of both bodies; two bodies make a demo, not a test set.
X-ray microtomography of entire mouse brain (to CCFv3)3.3 teravoxel micro-CT of mouse brain vs Allen CCFv31 brainLandmark pairs used in non-rigid registration (Humbel, Tanner et al.)both imagesdirectCC BY 4.0openleft out
one brain; its landmarks served the fit
Landmarks used in fitting, not held out
ZAPBench zebrafish whole-brain light-sheet (Google 2025)4D light-sheet calcium imaging of one larval zebrafish brain (~2 h): raw vs aligned frames, and functional frames vs a high-resolution anatomy stack; same subject7879 frames to a reference; functional to anatomyalignment flow fields from Google's pipeline (method output); segmentation of ~70k neurons and annotationsdirectCC BY 4.0openleft out
one fish; the frame alignment flows are the provider's
EM of the same fish is under proofreading, not public; Janelia 'Fish Fire&Wire' gives it by request with a publication embargo to Dec 2027.Apply known smooth deformations to aligned frames chunk by chunk for exact motion truth; check with segmentation overlap across frames

Remote sensing · 153

DatasetProblemPairsChecked againstAccessLicenceStatusNotesTo make the truth
3MOSOptical to SAR from 5-6 satellites (GF-3, ALOS, Sentinel-1, Radarsat, RCM), 8 scene types, same scene113,000 pairs 256x256 (README; arXiv v1 said 155K)Expert-selected control points, affine with RANSAC, error <1 px; eval adds random rotation +-5 deg and shift +-30 pxboth imagesdirectThe 3MOS Dataset is released under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License [CC BY-NC-ND 4.0].non-commercialtruth to make
patches aligned by experts' control points (affine, under 1 px), scored only through added offsets: a smooth known warp gives a synthetic SAR-optical truth over five SAR sensors; CC BY-NC-ND, 256 px patches, OneDrive; after Major TOM
Truth is synthetic offset on expert-aligned pairs; affine only. status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Remote sensing registration review 2026).Two people click 20 independent check points on 50 test pairs to measure base alignment
Aerial NIR-RGB set (Lu et al., CoMIR evaluation)Aerial near-infrared to RGB patches (likely Zurich QuickBird images, unverified)864 pairs 300x300Known synthetic rigid transform applied to co-registered NIR/RGBboth imagesdirectcc-by-4.0opentruth to make
co-registered aerial NIR/RGB patches with rigid offsets only (CC BY 4.0, Zenodo, 229 MB): a smooth known warp gives an exact multispectral truth on 300 px patches; low
Status change: the existing MultiRegEval entry (zenodo 4587903) lists only cell and tissue data; this record adds the aerial part.Add smooth non-rigid warps to the same co-registered patches
AerialMegaDepthpseudo-synthetic aerial renders (Google Earth 3D) registered with ground photos (MegaDepth) of landmarksmany aerial-ground pairs with depth (count in paper)camera poses and depth from joint reconstruction; renders aligned by constructionboth imagesdirectnot stated; Google Earth renders subject to Google terms (unverified)unknownwaits for a tool
ground photos against aerial renders of Google Earth 3D: camera poses and depth, a 2D/3D task; renders under Google's terms; waits for 2D/3D pose error
Aerial-to-ground, not satellite; truth from 3D reconstruction.
AerialVLUAV RGB frames (gimbal camera) vs satellite map tiles, ~70 km of trajectories, varied terrain, heights, lighting18,361 aerial images with 14,096 map patches; 11 sequences 3.7-11 kmGNSS positions (NovAtel, 1.5 m RMS) and attitude; DEM released 2026one imagedirectMIT (repository); data terms not separate (unverified)opentruth to make
UAV frames against map tiles with GNSS poses only: people would place 10 points per frame pair; one region, map source unclear; low (CRASAR-U-DROIDs covers drone to satellite)
Position truth only.Project frame footprint with pose + DEM, then refine with manual points on a subset
ALTO (Aerial-view Large-scale Terrain-Oriented)helicopter/UAV nadir images vs satellite/ortho reference, large terrain areas (US)GPR competition rounds 1-2 (ICRA 2022); full set 'coming soon'GPS posesone imagedirectBSD-3-Clause (repository)openleft out
its images or its truth are not released: only two place-recognition rounds out, the full set never released
Retrieval-oriented; full dataset never released per README.Manual point pairs on a subset
Apollo 17 LROC NAC controlled orthomosaic and Hasselblad registrationsTaurus-Littrow valley: NAC orthomosaic 0.5 m on 1.5 m NAC DEM; Apollo surface photos registered to the orbital imageorthomosaic vs individual NAC images; surface panoramas vs orbital mapcontrolled to ALSEP central station coordinates (~10 m); DEM vs LOLA sigma 1.3 m; station and instrument positions located by joint analysis (Haase et al. 2018)both imagesdirectUSGS/PDS public (unverified)opentruth to make
one site: the NAC images and the 0.5 m mosaic share one control (provider), so only a known warp; low
Human-placed hardware (ALSEP, LRV tracks) are physical markers visible from orbit.
ASTER L1T (VNIR/SWIR/TIR from one platform)Terra ASTER VNIR 15 m, SWIR 30 m, TIR 90 m, plus 3N/3B along-track stereo, 2000-nowVNIR vs TIR (optical-thermal) and VNIR vs SWIR of the same scene; nadir vs backward stereosame-platform geometric model; known inter-subsystem registration offsets corrected in L1T (accuracy per product, unverified)one imageaccountNASA/METI open (no restrictions, unverified)openneeds an account
the download needs an account
Primary page not confirmed in this sweep; classic optical-to-thermal satellite pair from one platform.Known warp between VNIR and TIR for exact optical-thermal truth
ATR-UMMIM / ATR-UMMIR (UAV visible-infrared registration benchmark)UAV visible (1920x1080) to thermal infrared (640x512), DJI H20T/H20N, 80-300 m, day and night7,969 triplets (raw VIS, raw IR, registered VIS)Semi-automatic: manual time sync, expert coarse warp, automatic refinement; pixel-level aligned VISboth imagesrequestNot statedunknownneeds a request
the data needs a request or a data use agreement
Aerial, not satellite; download via Baidu or by contacting the authors.Two people click 10 points on 100 test triplets to measure residual error
Caltech Aerial RGB-Thermal Dataset in the Wildsynchronised RGB and thermal video from aircraft/drone over rivers, coasts, deserts, forests across the US, with GPS and IMUsynchronised RGB-thermal frame pairs (many trajectories)none for registration (synchronised, different lenses); GPS/IMU poses; thermal semantic labelsone imagedirectsee CaltechDATA record (NOASSERTION on GitHub, unverified)unknowntruth to make
synchronised airborne RGB-thermal frames without registration truth: people points or a calibrated homography per frame; video, licence unclear (NOASSERTION); low
Natural scenes with thermal inversion; RGB-T alignment is left to the user.Calibrate RGB-thermal extrinsics with a heated target, then homography per frame at altitude gives near-exact truth; manual points on 50 pairs
Capella SAR Open DataX-band SAR spotlight/stripmap, sub-metre, worldwide samples and eventsSAR vs optical; some repeat acquisitionsnone (geocoded products)one imagedirectCC BY 4.0opentruth to make
open VHR SAR (CC BY 4.0, anonymous S3) against open optical (NAIP, Maxar ODP): two people placing 20+ tie points per pair as SpaceNet 9 did; second source to Umbra for a ten-site optical-SAR set
TileDB copies support chunked reads natively.Check points vs open optical by two operators
CDD (Lebedev seasonal change detection)Google Earth seasonal pairs, 0.03-1 m, 11 original pairs7 pairs 4725x2700 + 4 smaller; 16,000 crops 256x256aligned by Google Earth; change labels ignoring seasonal changeone imagedirectunverifiedunknowntruth to make
Google Earth seasonal pairs aligned by the provider, licence unverified: known warp or people points; low
Strong seasonal appearance change, useful for robustness of similarity metrics.Known warps
Chang'e-2 / Chandrayaan-2 lunar mapping productsChang'e-2 7 m global DOM/DEM (CLEP); Chandrayaan-2 OHRC 0.25 m and TMC-2 5 m (ISRO)vs LROC NAC/WAC, Kaguya TC, LOLAmission control networks (accuracy per product, unverified)one imageaccountCNSA/CLEP and ISRO/ISSDC terms, registration required (unverified)unknownneeds an account
the download needs an account
Listed for completeness; accounts needed, terms not checked.Register to LOLA/SLDEM hillshade; Robbins craters as landmarks
CityMMReg-dataUrban SAR-RGB (GF-3/GF-2; Capella vs Google Earth), IR-RGB (WorldView-2 vs Esri), SAR-IR (SpaceNet 6)4 test groups of 236-300 px tiles (counts not checked)Groups 2-3: 15 sets of manual control points; groups 1 and 4: random known transforms (.mat)both imagesdirectNot statedunknowntruth to make
the public repository holds 4 point-annotated tile pairs (3 Capella-Google Earth, 1 WV-2-Esri; ENVI GCP files of about 20 points) of 256-300 px, the rest random transforms; no licence; low
Mix of real cross-date pairs with manual points and synthetic ones.More annotators on the control-point groups; 20 points per pair
CM-Bench infrared-satellite geo-localisation setaerial infrared images vs satellite imagery for geo-localisation (plus ground IR-VIS sets in the benchmark)count not seen (unverified)manually annotated ground-truth correspondencesboth imageshiddenunknownunknownleft out
its images or its truth are not released
Release pending at time of paper.
Copernicus DEM GLO-30 / GLO-90 (COG)global DSM at 30 m and 90 m (2021 release)DEM/hillshade vs optical or SAR (DEM-to-image registration); DEM vs SRTM/ALOSnone per pair; DEM used for orthorectification of Sentinel-2, so aligned with refined S2 by construction (to DEM accuracy)one imagedirect'GLO-30 Public and GLO-90 are available on a free basis for the general public under the terms and conditions of the Licence' (Copernicus DEM licence)openleft out
no second image: a DEM, kept as the terrain reference for making truth
Needed for terrain-aware registration and for simulating SAR geometry.Hillshade rendered with the image's sun angle vs the image: aligned by construction up to DEM error
CORE3D public dataWorldView-2/3 PAN+MS multi-date images over Jacksonville, Omaha, UCSD and other US sites, with airborne LiDAR and building labels26 (Jacksonville), 43 (Omaha), 35 (UCSD), 44 (another site) WV-3 images; any pair per siteairborne LiDAR and semantic labels; GCPs for Jacksonville used by SatDepth come from a separate sourceone imagedirectpublic release by IARPA; imagery courtesy DigitalGlobe; citation required (exact licence text not on page)unknownleft out
no image twice: SatDepth holds these WorldView scenes with dense truth
Base data for several satellite matching sets.RPC+LiDAR projection for dense truth; manual check points on 10 pairs
CRASAR-U-DROIDs (disaster orthomosaics with building-polygon alignment adjustments)sUAS, crewed-aircraft and satellite orthomosaics (+ some UAS DSMs) of 10 US disasters vs Microsoft building footprints; same buildings seen in several orthomosaics265 orthomosaics, 21,716 buildings, 122,502 building views; 70.6 gigapixels over 68 km2human-made translational alignment adjustments (vectors) moving polygons onto the imagery, per orthomosaic; building_id links the same building across orthomosaicsone imagedirect: Hugging Face dataset, not gated, one GeoTIFF a file (219 GB in all)CC BY 4.0 (dataset card)openin the corpuspriority 1
The Hugging Face repository is ungated and CC BY 4.0. Each satellite or crewed-aircraft crop is cut to a sUAS orthomosaic's boundary: 38 sUAS areas in five hurricanes also have one or more such crops, sharing tens to thousands of building_ids (statistics.csv), and each orthomosaic has its own human-drawn translational adjustments that move the footprints onto it. A building's adjusted centroid in both images gives point pairs drawn by people on both, but the sUAS files are large (median 2.1 GB, up to 18 GB), so the build waits for the large-image path; whether a non-rigid residual is left after the corpus's alignment must be measured.
Gigapixel UAS orthomosaics are a real out-of-core case; cross-source (UAS vs satellite vs crewed) views of the same building give multi-sensor correspondences.Use adjusted polygon corners of the same building_id in two orthomosaics as point pairs; check 10 orthomosaic pairs by a second annotator
CrossLoc benchmark (Urbanscape, Naturescape)drone photos (2D) against swisstopo 3D terrain and building data with synthetic renders (3D), Lausanne area2 scenes, real in-place and out-of-place drone trajectories plus large synthetic setsreal image poses from RTK geo-tags and photogrammetric adjustment; synthetic poses exactboth imagesdirectrepository MIT; data terms not read (unverified)unknownwaits for a tool
drone photos against swisstopo 3D terrain renders: pose truth; waits for 2D/3D pose error
real and synthetic pairs from the same place make a sim-to-real test
CTX controlled targeted DTMs and orthoimages (USGS, ASP)Mars CTX stereo pairs (~6 m) globally bundle-adjusted (ISIS jigsaw) then ASP DTMs + orthoimages, with MOLA shot comparisonsthousands of CTX pairs co-located with HiRISE targetsrelative global control; MOLA shots CSV and intersection-error rasters per DTM; qa_metrics.txtone imagedirectCC0 1.0opentruth to make
CTX orthos co-located with HiRISE on one control: a known warp across 6 m and 0.25 m; after HiRISE
CTX orthos co-located with HiRISE orthos give a cross-resolution (6 m vs 0.25 m) pair on a common control.CTX ortho vs HiRISE controlled ortho: same frame, known scale; MOLA shots as absolute check
DDHR dataset (DDHRNet, GF-2/GF-3 Xi'an, Pohang, Shandong)GaoFen-2 optical to GaoFen-3 SAR, high resolution, three regionstiles per region (count in paper, unverified); large-resolution scenes announcedco-registered for land-cover labels by authors (method in paper, unverified)one imagerequestnot stated; README says contact author because of data governanceaskneeds a request
the data needs a request or a data use agreement
Status unclear: links present but 'contact if necessary'.Check points on scene pairs by two operators
Declassified CORONA / Hexagon KH-9 imagery (USGS)1960s-1980s film reconnaissance frames (2-10 m) vs modern orthophotos/Sentinel-2any historical frame vs modern referencenone (frames uncontrolled; users place GCPs)one imageaccountUS public domain (USGS EarthExplorer, unverified)openneeds an account
the download needs an account
Primary page not confirmed in this sweep; extreme temporal gap and panoramic distortion.Two operators click 30 GCPs on 5 frames vs modern ortho + Copernicus DEM
DFC2013 Houston (hyperspectral + LiDAR)CASI hyperspectral 144 bands vs LiDAR DSM, 2.5 m, University of Houston (2012)one 349x1905 scene pairco-registered by provider; 15-class labelsone imageaccountunverifiedunknownneeds an account
the download needs an account
Classic HSI-LiDAR pair; primary page not confirmed in this sweep (IEEE DataPort / UH site).Known warp for exact HSI-to-DSM truth
DFC2014 (thermal infrared hyperspectral + VIS, Thetford Mines)airborne LWIR hyperspectral (84 bands, 7.8-11.5 um, ~1 m) and RGB (20 cm) from separate flight lines, same platform, collected simultaneously 21 May 2013, urban QuebecLWIR mosaic vs RGB mosaic; 74 individual orthorectified LWIR hypercubes (paper contest) vs RGB sub-imagesboth georeferenced by provider (Telops); no check points; classification labelsone imagedirectnot stated on page (download kept available, unverified)unknowntruth to make
airborne LWIR hypercubes and RGB from separate flight lines, one town, provider georeferencing: people points on sub-scenes; licence not stated; low
Rare thermal-hyperspectral vs VHR visible aerial pair; separate flight lines imply residual offsets.Building corners and road markings clicked in LWIR band image and RGB by two operators on 10 sub-scenes
DFC2015 (Zeebrugge LiDAR + 5 cm orthophoto)airborne LiDAR (65 pts/m2, DSM) and 5 cm RGB orthophoto collected simultaneously, harbour and urban area, Belgium, 13 Mar 20117 tiles: 10,000 x 10,000 px ortho + up to 5000x5000 DSM + point cloudsimultaneous acquisition, both georeferenced to WGS-84 (co-registered by construction); semantic labels laterone imagedirectnot stated on page (Royal Military Academy data, unverified)unknowntruth to make
5 cm ortho and LiDAR DSM of one harbour, aligned by construction: known warp; licence not stated; after ISPRS Potsdam
Image-to-DSM pairs at 5 cm GSD, a big-image multimodal case.Known smooth warp between DSM and ortho (exact); roof corners as check points
DFC2016 (Vancouver: DEIMOS-2 L1B/L1C two dates + ISS Iris video)DEIMOS-2 pan 1 m and MS 4 m on 31 Mar and 30 May 2015 as Level 1B (raw geometry + RPC) and Level 1C (manually orthorectified); ISS Iris UHD video 3 fps, frames ortho-rectified to 1 m (3840x2160)2 dates x (pan, MS) x (1B, 1C); video frames vs DEIMOS-2Level 1C 'cover exactly the same ground area for both dates' after manual orthorectification; 1B has RPCs onlyone imagedirectnot stated on page (Deimos Imaging / UrtheCast data for the contest, unverified)unknowntruth to make
one pair of dates (manually orthorectified L1C) plus L1B with RPC: people points, one scene only; licence not stated; low
Raw L1B + RPC alongside manually orthorectified L1C is exactly the raw-vs-reference setup for registration.Two operators click 20 points on the two L1C dates to measure the manual orthorectification residual; L1B-to-L1C mapping via RPC is the reference transform
DFC2017 (Local Climate Zones, Landsat + Sentinel-2 + OSM)Landsat-8 multi-date and Sentinel-2 resampled to 100 m, OSM rasters and vectors, 9 citiesmulti-date Landsat vs Sentinel-2 vs OSM per citysuperimposable rasters by construction; LCZ labelsone imageaccountopen access on IEEE DataPort (Copernicus and USGS data; OSM ODbL)openneeds an account
the download needs an account
100 m resampling makes it weak for registration; listed for completeness.Use the linked full-resolution Sentinel-2 with OSM for map-to-image tests
DFC2018 Houston (multispectral LiDAR + hyperspectral + VHR RGB)Optech Titan 3-wavelength LiDAR (DSM/intensity 0.5 m), CASI hyperspectral 48 bands 1 m, DiMAC RGB 5 cm; University of Houston, one flight 16 Feb 20173 modalities per tile over the same areaco-registered by provider (same aircraft, georeferenced); 20-class land-cover labelsone imageaccountCreative Commons Attributionopenneeds an account
the download needs an account
Same-platform multi-sensor: registration truth by construction, label maps as extra check.Known warp between hyperspectral, LiDAR DSM and RGB gives exact multimodal truth at three GSDs
DFC2021 (DSE: S1/S2/Landsat/VIIRS night lights; MSD: NAIP 2013/2017 + Landsat)DSE: Sentinel-1, Sentinel-2, Landsat-8 and VIIRS DNB night lights (750/500 m) all resampled to 10 m, 60 training tiles; MSD: Maryland NAIP 1 m 2013 and 2017 + Landsat-8 5 dates + NLCDday optical vs night lights vs SAR per tile (DSE); NAIP 2013 vs 2017 (MSD)co-located by resampling to a common grid; weak labelsone imageaccountnot seen (IEEE DataPort, unverified)unknownneeds an account
the download needs an account
Only public contest data pairing VIIRS night lights with optical and SAR on one grid; MSD NAIP pair is state-wide.Known warps; NAIP 2013 vs 2017 road-junction check points on 10 tiles
DFC2023 (optical + SAR building roof types, 17 cities)SuperView-1 (0.5 m) / GaoFen-2 (0.8 m) optical with GaoFen-3 SAR (1 m) tiles, plus nDSM from GaoFen-7/WorldView stereo (2 m)optical-SAR tile pairs, 17 cities on 6 continentsco-registered for the contest (method not stated); building polygons and nDSMone imagedirect'data remains openly available to the community' (exact licence not seen)unknowntruth to make
optical-SAR-nDSM tiles of 17 cities co-registered for the contest (little room): known warp; exact licence not seen; low
Optical-SAR-nDSM triplets: the nDSM can explain SAR layover offsets.Building footprint corners in optical vs SAR by two operators on 30 tiles; layover predicted from nDSM
DFC2024 (flood mapping: Sentinel-1 and HLS)Sentinel-1 C-band SAR 10 m (track 1) and HLS 30 m (track 2) with Copernicus DEM, MERIT DEM, GSW occurrence, WorldCoverSAR/optical vs DEM vs land cover per flood eventco-located grids; water labels from CEMS, OPERA DSWx CalVal, CERFACS simulationsone imagedirectnot stated on page (unverified)unknowntruth to make
same-grid flood stacks at 10-30 m: known warp; low
Registration not part of the task; same-grid multimodal stacks only.Known warps
DFC2026 (SAR temporal storytelling: Capella InSAR-compatible stacks)Capella X-band SAR: ~1,582 collects in coordinated stacks over one year (Spotlight, Sliding Spotlight, Stripmap; SLC, GEO, GEC, some CPHD), varying incidence and look direction>17,000 possible pairs listed in a per-pair CSV (some AOIs > 50 images)none for registration; SLC pairs 'InSAR compatible where repeatability permits'one imagedirectopen data on public S3 without authentication; contest T&C apply; licence not named on page (Capella open data is CC BY 4.0 per AWS registry)openleft out
no image twice: these are Capella open-data stacks (the Capella row)
Largest open VHR SAR repeat-pass stack set: SAR-to-SAR co-registration is a hidden prerequisite for every pair.InSAR coherence after co-registration as an image-based quality score; corner reflectors / persistent scatterers as point truth
Drone-satellite retrieval sets (University-1652, SUES-200, DenseUAV, GTA-UAV): fail pixel criteriondrone (real or rendered) images vs satellite tiles of the same locationthousands of drone-satellite pairs per setlocation/building ID or GPS (GTA-UAV adds exact 6-DoF poses from the game engine)one imagedirectDenseUAV and GTA-UAV repos Apache-2.0; University-1652 and SUES-200 terms not checkedunknowntruth to make
drone or rendered views against satellite tiles with location IDs only: people points per pair; terms unchecked for two of the four sets; low
Retrieval benchmarks, no pixel correspondences; GTA-UAV poses could be turned into dense truth with the game's terrain (synthetic). status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Remote sensing registration review 2026).For GTA-UAV render dense correspondences from pose + terrain; others: manual points on a subset
DroneVehicle (RGB-infrared UAV)UAV RGB vs thermal infrared, urban roads and parking, day to night28,439 RGB-IR pairs 640x512none for registration; vehicles annotated separately in RGB and IR with oriented boxes (e.g. 389,779 cars in RGB, 428,086 in IR)both imagesdirectnot stated (unverified)unknownneeds an account
Baidu only (the Google Drive link is still 'to be released'); per-modality vehicle boxes on near-aligned 640x512 frames
Pairs are roughly aligned but not perfectly; separate per-modality boxes give paired object centres as weak correspondences.Box centres of the same vehicle in RGB and IR as point pairs; manual 10-point check on 50 pairs
DSIFN-CDGoogle Earth bitemporal pairs over 6 Chinese cities6 large pairs; 394 crops 512x512 (+48 test)aligned by Google Earth; change labelsone imagedirectnot stated (Google Earth terms)unknowntruth to make
Google Earth pairs aligned by the provider, change labels only: known warp; Google terms; low
Large source pairs available.Known warps
DynamicEarthNetdaily PlanetScope (3 m) time series 2018-2019 over 75 AOIs + Sentinel-1/275 AOIs x ~730 days, 1024x1024Planet analysis-ready co-registration (residual unverified); monthly land-cover labelsone imagedirectunverifiedunknowntruth to make
daily PlanetScope cubes co-registered by Planet: known warp between days; licence unverified, hundreds of GB; low
Primary page not confirmed in this sweep. URL is the dataset's GitHub (homepage links from there).Known warps between days; check points on 10 AOIs
EarthMatch / astronaut photo localisation data (EarthLoc, AIMS)handheld ISS astronaut photos (oblique, varied focal length, clouds) to Sentinel-2-like satellite basemap tiles268 astronaut photos x top-10 candidate tilesphoto footprint from NASA's manual localisation records (centre point / footprint), not pixel correspondencesone imagedirectnot stated (unverified)unknowntruth to make
astronaut photos with footprint centres only: people would place 10-15 points per photo; licence not stated; low
Extreme viewpoint and scale; truth too coarse for pixel registration scoring.Hand-place 10-15 points per photo on a 50-photo subset; homography from them
EOPF Sentinel Zarr Samples (Sentinel-1/2/3 in Zarr)Sentinel-2 L1C/L2A tiles, Sentinel-1 SLC (2023-2025) and Sentinel-3 in EOPF Zarrany two dates of a tile (rolling ~1-year archive for S2 L2A over Europe; L1C back to 2015 in STAC)same as Sentinel-2 products (GRI refinement); none per pairone imagedirectpublic, open and free (CDSE service page); Copernicus termsopenleft out
no image twice: the same Sentinel-2 products as the Sentinel-2 row, in Zarr (useful later as a native out-of-core format)
Native Zarr makes it the natural out-of-core test format; operational Zarr production expected late 2026/early 2027.Known smooth warp applied chunk-wise to a full Zarr tile; two refined dates as real pair
EuroSDR TIME benchmark (historical aerial images)digitised historical aerial image blocks and time series since the 1950s from national mapping agencies (AT, CY, EE, FI, NO, SI), the Italian aerial photo library and Warsaw UTmulti-epoch blocks per site (counts in the paper/table, not read)ancillary flight/camera data and ground control points supplied per datasetboth imagesrequest'available for research purposes' (by email to 3dom@fbk.eu)non-commercialneeds a request
the data needs a request or a data use agreement
Request by email, so not checked directly.Second operator re-measures GCPs on 10 images per epoch; modern ortho as reference
Galileo Europa controlled observations and mosaics (USGS)481 Galileo SSI images of Europa with photogrammetrically improved positions; 92 observation mosaicsoverlapping images and image vs mosaicphotogrammetric control network (improved relative locations)one imagedirectCC0 1.0opentruth to make
overlapping Galileo frames on a photogrammetric network whose tie points are not shipped: people points or a known warp; low
Small, low-texture, varied-illumination planetary set.Control-network tie points (if released with the product) as truth; else manual points on 10 overlaps
Google Earth multi-temporal aerial matching set (Park et al.)Google Earth images of the same places in 2015, 2017, 2019 (Landsat-7/8, WorldView, QuickBird sources), large viewpoint changes~9,000 images; DVF-NET used 2,000 train / 500 testsynthetic affine/perspective transforms on top of roughly aligned datesboth imagesdirectGoogle Earth terms (unverified)unknownleft out
its images or its truth are not released: no download page found
Download page not found; cited as 'Google Earth dataset proposed by Park et al.'.Manual check points on 50 real pairs
GoogleEarth and GoogleMap homography sets (DLKFM, CVPR 2021)Cross-season Google Earth (Boston 2018 vs 2019) and Google map vs satellite, same place~8k train each; 850 (Earth) and 888 (Map) test pairs, 192x192 inputsSynthetic homography (random corner perturbation) on tiles aligned by map coordinatesboth imagesdirectNot statedunknowntruth to make
192 px Google Earth/Maps tiles with synthetic homographies, no licence: a smooth known warp; low
Low resolution; truth synthetic. Google imagery terms apply.Add smooth local warps; check base tile alignment on 20 pairs
GUSO (global ultra-high-resolution SAR-optical)Ultra-high-resolution SAR (ICEYE, Capella, Umbra open data, 0.16-0.98 m) to optical, 319 cities in 78 countries593,335 pairs; urban, rural, plain, hill, water, disaster splitsFine-grained manual registration of each pair (per README); evaluation protocol not checkedboth imagesrequestThe GUSO dataset can be used for academic purposes only and need to cite the following paper, but any commercial use is prohibited.non-commercialneeds a request
the data needs a request or a data use agreement
Google form before download links. New 2026, very large.Two people click 20 check points on 50 test pairs per scene type
Harmonized Landsat Sentinel-2 (HLS v2.0)Landsat 8/9 OLI and Sentinel-2 MSI surface reflectance on a common 30 m MGRS grid, 2-3 day revisitLandsat-Sentinel cross-sensor pairs and multi-date pairs per tileLandsat registered to the Sentinel-2 GRI-based reference in v2.0 (per HLS documentation, unverified); co-registered by constructionone imageaccountCreative Commons BY 4.0openneeds an account
the download needs an account
Cross-sensor stacks with documented harmonisation; account needed.Known warp between L30 and S30 of the same day for exact cross-sensor truth
Hessigheim 3D (H3D) UAV LiDAR + imagery, 4 epochsUAV LiDAR (800 pts/m2) and 2-3 cm imagery from the same platform; textured meshes; epochs Mar 2018, Nov 2018, Mar 2019, plus 2016 manned ALSmulti-epoch point clouds and meshes of one village area; image-LiDAR within an epochsame-platform co-registration within epoch; epochs georeferenced (registration accuracy in paper, unverified); 11-class labelsone imagerequestbenchmark terms on ifp site (registration form, unverified)unknownneeds a request
the data needs a request or a data use agreement
Multi-epoch 3D change and registration testbed; images may need a request.Known rigid/non-rigid warp between epochs; stable building corners as check points
Hi-UCD (Estonia urban change)Estonian Land Board aerial orthophotos 0.1 m, multi-year (2017-2019)1,293 pairs 1024x1024 (unverified)orthophotos from national programme (GCP-based); semantic change labelsone imagedirectunverified (Estonian orthophotos are open data)unknowntruth to make
Estonian orthophoto change pairs, page not confirmed: known warp, or the national orthos at the source; low
Page not confirmed in this sweep; Estonian open orthophotos also usable directly as large scenes. URL is the dataset's GitHub (homepage links from there).Known warps; national check points
HiRISE controlled DTMs and orthoimages (USGS ARD on AWS)Mars HiRISE stereo pairs: 1 m DTM + 0.25-0.5 m orthoimages of both images of each pair, controlled to MOLAhundreds of released DTM sites; each gives 2 orthoimages of the same terrain from two dates/anglesboth orthos rectified on the same DTM in a MOLA-controlled frame: co-registered by construction to sub-metre level (residual unverified)one imagedirectCC0 1.0opentruth to make
the two orthos of a HiRISE stereo pair sit on one controlled DTM (CC0, anonymous COG, 1.3 GB each): a known warp applied by blocks gives an exact truth on gigapixel pairs, the large-image case; first synthetic
Ready-made real large pairs: two orthos from different orbits differ in illumination and season but share geometry; ideal for out-of-core exact-ish tests.Apply a known smooth warp to one ortho chunk by chunk for exact truth; or pair the uncontrolled RDR (a-priori SPICE) with the controlled ortho for a real misregistration with a known reference
HiRISE uncontrolled RDR observations (USGS ARD COGs)Mars HiRISE red and colour RDRs map-projected with a-priori SPICE pointing (not controlled)any RDR vs controlled ortho/CTX of the same area; repeat RDRs for changenone (explicitly 'will not co-register with controlled data products')one imagedirectCC0 1.0opentruth to make
uncontrolled RDR against the controlled ortho of the same observation: the provider's control solution as truth (provider); after HiRISE controlled
Natural misregistered input whose controlled counterpart exists where a DTM was made.Controlled ortho of the same observation = reference; derive the RDR-to-ortho transform from the DTM product as truth
Hong et al. multimodal benchmarks (HS-MS Houston2013, HS-SAR Berlin, HS-SAR-DSM Augsburg)hyperspectral vs multispectral (Houston2013, simulated MS), hyperspectral (EnMAP-simulated HyMap) vs Sentinel-1 SAR (Berlin), HySpex hyperspectral vs Sentinel-1 SAR vs DSM (Augsburg)3 scenes, each a co-registered multimodal stackco-registered by the authors on a common grid; land-cover labelsone imagedirectcode GPL-3.0; datasets 'freely and openly available' (data licence not stated)opentruth to make
three small co-registered HS/MS/SAR/DSM stacks: known warp; data licence not stated; low
Small but rare HSI-SAR-DSM stacks; Trento and Chikusei not confirmed in this sweep.Known warp across modalities for exact HSI-SAR / HSI-DSM truth
HRSC MC-30 controlled mosaics and multi-orbit DTMs (Mars Express)Mars Express HRSC nadir/stereo/colour, 12.5 m orthomosaics and 50 m DTMs per MC-30 quadrangle, controlled to MOLAHRSC vs CTX/HiRISE/THEMIS; overlapping HRSC orbitsmulti-orbit block adjustment tied to MOLA (Gwinner et al.); stated accuracies in papers (unverified)one imagedirectESA/DLR/FU Berlin terms (CC BY-SA 3.0 IGO for images, unverified)unknowntruth to make
MOLA-controlled HRSC mosaics (page not confirmed, ESA/DLR terms): known warp; low
Primary page not confirmed in this sweep.HRSC vs CTX mosaic tie points; MOLA shots as absolute check
HRSCD (High Resolution Semantic Change Detection)IGN BD ORTHO aerial RGB 0.5 m, France (Rennes, Caen), 2006 vs 2012291 co-registered pairs 10,000 x 10,000 pxcoregistered orthophotos (IGN); change and land-cover maps from Urban Atlasone imagedirectCC BY-NC-SA (dataset); images under IGN 'licence ouverte'; labels from Copernicus Urban Atlas termsnon-commercialneeds an account
IEEE DataPort; co-registered IGN orthos with change labels, no registration truth
100-megapixel orthophoto pairs: good mid-size out-of-core case; IGN orthos have national GCP-based accuracy.IGN geodetic points / building corners as check points; or known warp chunk by chunk
Hyperspectral registration repository (CiTIUS)AVIRIS hyperspectral, same site at two dates (Jasper Ridge, Santa Barbara, Baraboo Hills, Crown Point); plus single scenes5 multi-temporal pairs; 4 single scenes for synthetic testsReference similarity parameters (scale, rotation, shift) per pair; how derived not statedboth imagesdirectCreative Commons Attribution Non Commercial Share Alike 3.0 Unportednon-commercialtruth to make
5 real multi-temporal AVIRIS pairs with similarity parameters of unstated origin: people points; CC BY-NC-SA; low
Real multi-temporal pairs but only similarity truth.Two people click 20 points per pair to check the reference parameters
IARPA MVS3DM (Multi-View Stereo 3D Mapping, San Fernando)50 WorldView-3 PAN/MS/SWIR images over 100 km2 near San Fernando, Argentina, 2014-2016 multi-dateany image pair of the 50 (1,225 pairs)20 cm airborne LiDAR DSM on 20 km2 (geometry truth); correspondences derivable through RPC + DSMone imagedirect'made publicly available with no restrictions to support continued research' (imagery courtesy DigitalGlobe)openleft out
no image twice: SatDepth holds these scenes (its Argentina AOI) with dense truth
Full-resolution NITF scenes are a genuine large-image case; RPC bias between dates gives real misregistration to correct.Project LiDAR DSM through refined RPCs to get dense truth for any pair (the SatDepth recipe); surveyed GCPs not included
ICEYE SAR Open DatasetX-band SAR curated selection, various modesSAR vs optical; some time seriesnoneone imagedirectCreative Commons License CC BY 4.0opentruth to make
open X-band SAR (CC BY 4.0): same people-point recipe as Umbra and Capella, fewer scenes; low
Check points vs open optical
IGN BD ORTHO and BD ORTHO Historique (France)French national orthophotos 20 cm (recent) and 50 cm, plus historical orthos (1950-1965) of the whole countryhistorical vs modern ortho of the same commune (60-70 year gap); successive campaignsorthorectified by IGN on national geodetic control (accuracy per product spec, unverified)one imagedirectLicence Ouverte / Etalab (IGN open data, per HRSCD page)opentruth to make
national orthos 1950-65 and today (Licence Ouverte): two people placing 20 points on 10 communes give an expert truth across 60 years; heavy JP2 mosaics
Primary page not confirmed in this sweep; historical-to-modern pairs are a strong temporal-gap case.Road junctions / building corners as check points by two operators on 10 communes; known warp for exact truth
Image Matching Challenge 2024 (aerial and aerial-ground scenes)image collections incl. aerial (drone) and aerial-to-ground scenes, plus temporal-change and historical scenesper-scene image sets (hidden test)camera poses from SfM reconstructions; metric mAA on camera translation after registrationboth imageshiddenKaggle competition terms (unverified)unknownleft out
its images or its truth are not released
Test data hidden; 3D pose truth rather than 2D registration.
Inria/Girard cadaster map alignment data (mapalignment)OSM / cadaster building polygons vs aerial orthophotos (Inria Aerial Image Labeling and Bradbury buildings sets, US/EU cities)city tiles; per-tile polygon sets (count in papers)aligned polygons as reference; misalignment simulated by random smooth displacement fields (exact truth); real OSM offsets corrected by noisy supervisionone imagedirectcode repo licence not checked; Inria Aerial dataset has its own terms (unverified)unknownleft out
no second image: building polygons against an image, with synthetic displacements
Classic map-to-image alignment setup with synthetic smooth deformations, i.e. the 'known warp' recipe applied to vector data.Smooth random displacement applied to polygons is exact truth; real offsets need manual correction on a test subset
ISPRS Vaihingen and Potsdam 2D semantic labelling (orthophoto + DSM)Aerial true orthophoto (IRRG or RGBIR) and DSM from dense matching, same grid, urban GermanyVaihingen 33 patches (9 cm); Potsdam 38 tiles 6000x6000 (5 cm)None for registration: TOP and DSM aligned by constructionone imagedirectNot stated on the pages seenunknowntruth to make
true ortho and DSM of one block, aligned by construction: a known warp gives exact optical-to-height truth; licence and download terms to confirm
Fails truth as is; ideal source for exact optical-height (LiDAR-like) truth.Apply known smooth deformations to the DSM against the TOP: exact multimodal dense truth
ISPRS WG I/4 satellite stereo benchmark (Catalonia: Terrassa, La Mora, Vacarisses)Cartosat-1 stereo (2.5 m), WorldView-1 stereo (0.5/0.76 m, 10,000x10,000 L1B with RPC), Pléiades triplet (2013), vs ICC 50 cm orthophotos and airborne LiDARstereo/triplet pairs per site; satellite vs orthoRPC block adjustment with GCPs from ICC orthos and LiDAR; LiDAR reference; epipolar-rectified versions added (2014)both imagesdirectacknowledge Euromap, DigitalGlobe, ESA/JAXA, ICC; registration by e-mail expected; FTP credentials published on the pageaskneeds a request
the page asks for registration by e-mail (FTP credentials posted); L1B stereo with RPC whose GCP truth is 3D
Older but has raw L1B + RPC + GCP-grade references: real geometry registration case.Use the ICC ortho as reference for WV-1 L1B; GCP list as check points
ISPRS/EuroSDR Multi-Platform Photogrammetry benchmark (Dortmund, Zeche Zollern, Zurich)Nadir and oblique aerial (PentaCam), UAV and terrestrial images of Dortmund centre and Zeche ZollernHundreds of images per block; image orientation taskGCPs by static GNSS (1 cm) and total-station facade targets; GCP and check-point image coordinates on all imagesboth imagesdirectThe data users should include the following acknowledgement ... "provision of the datasets by ISPRS and EuroSDR"openwaits for a tool
image orientation against surveyed GCPs (oblique, UAV, terrestrial); the page also has a registration form; waits for 2D/3D pose error
Geodetic truth; it is multi-view orientation rather than 2D pair registration. Page also mentions a registration form.
Kaguya/SELENE Terrain Camera scenes and USGS DTMs (ARD on AWS)Moon: Kaguya TC monoscopic, stereoscopic and spectral-support scenes (~10 m) as uncontrolled ARD; separate ASP DTMs with orthoimagesuncontrolled TC scene vs DTM-orthorectified TC; TC vs LROC WAC/NACnone for uncontrolled scenes; DTM products co-registered to their own orthos (LOLA alignment per product docs, unverified)one imagedirectCC0 1.0opentruth to make
uncontrolled TC scenes against DTM-orthorectified ones: provider truth or a known warp; low
Uncontrolled vs DTM-based versions give a natural misregistered pair with a reference.DTM ortho as reference; LOLA COPC points as absolute check
LEVIR-CD and LEVIR-CD+Google Earth VHR (0.5 m) bitemporal pairs, Texas cities, 5-14 years apart, 2002-2018637 pairs 1024x1024 (LEVIR-CD); 985 pairs (LEVIR-CD+)co-registered by Google Earth mosaicking (no residual reported); 31,333 changed buildings labelled and double-checkedone imagedirectuse of Google Earth images must respect Google Earth terms (page); dataset for research (unverified)non-commercialtruth to make
Google Earth pairs aligned by the provider, change labels only: known warp or people points; Google terms; low
Google Earth pairs carry parallax and small misregistration that hurts CD; no registration truth.Known smooth warp on one date (exact); unchanged building corners clicked on 30 pairs by two operators for real residuals
LEVIR-GM (LoRetta dataset)Multi-temporal optical satellite (2018-2022), 0.5-1024 m, 67 regions on six continents103,363 aligned pairs 1280x1280; 826,904 augmented pairs 512x512Known affine plus local non-rigid perturbations on aligned pairs: dense warp and matchability labelsboth imageshiddenNot statedunknownleft out
its images or its truth are not released
New Aug 2026; no download link seen. How base pairs were aligned not checked.Check base alignment with 20 expert points on 100 aligned pairs
LINZ New Zealand aerial imagery on AWS (nz-imagery), incl. historical surveysnational orthophoto collections of New Zealand: modern 5-7.5 cm urban and 0.2-0.4 m regional campaigns, plus historical survey orthos (1978-2007, 0.3-0.8 m); e.g. Christchurch 2011 (earthquake), 2015-16, 2018-19, 2020-21, 2021, 2023, 2025many epochs of the same city (decades apart); source of the WHU building CD pairorthorectified by LINZ/suppliers (accuracy in metadata, unverified); no per-pair truthone imagedirectCC-BY-4.0 (STAC collection metadata)opentruth to make
decades of open aerial orthos (CC BY 4.0, anonymous COG, 7.5 cm-0.8 m): two people placing 20 points per epoch pair in 10 towns give an expert truth on gigapixel pairs
Best open source found for decades-long, very high-resolution aerial pairs at gigapixel scale with an open licence.Two epochs of Christchurch at 7.5 cm with WHU building polygons as point pairs; LINZ geodetic marks as check points; known warp chunk by chunk
LoD-Loc datasets (UAVD4L-LoD, Swiss-EPFL)UAV photos (2D) against LoD2/LoD3 city building models (3D), Changsha and Lausanne2 areas (2.5 km2 and 8.18 km2), in-trajectory and out-of-trajectory query setsground-truth 6-DoF poses: RTK and mesh-based refinement for UAVD4L, CrossLoc geo-tagged poses for Swiss-EPFLboth imagesdirectcode MIT; data on Baidu Disk with no stated licence; Swiss LoD from swisstopo open dataunknownwaits for a tool
UAV photos against LoD city models: pose truth; waits for 2D/3D pose error
real UAV images against untextured wireframe models: a true image-to-CAD-city registration; Baidu Disk download may need an accountindependent check: survey 20 GCP-like building corners in a few images and compare with GT poses
LOLA altimetry (COPC point cloud) and SLDEM2015Moon geodetic reference: ~7 billion LOLA returns; SLDEM2015 merges LOLA with Kaguya TC DTMs (512 ppd)DEM/hillshade vs any lunar image (DEM-to-image registration)LOLA vertical precision ~10 cm, accuracy ~1 m (Mazarico et al.); defines the lunar control frameone imagedirectCC0 1.0 (LOLA COPC)openleft out
no second image: an altimetry point cloud
Not image pairs; the reference that makes truth for lunar pairs.Render hillshade at the image's sun geometry from SLDEM; register image to it; LOLA points as check
LROC NAC DTMs and orthophotos (RDR, incl. Apollo sites)Moon: LRO NAC stereo DTMs (2-5 m posts) with orthophotos of both stereo images at image and DTM resolution, hundreds of sitestwo orthos per DTM (different orbits/illumination); NAC ortho vs WAC/Kaguya TCSOCET SET triangulation tied to LOLA tracks; per-DTM LOLA RMS (e.g. Apollo 11: 1.81 m) and triangulation RMS in shapefileone imagedirectNASA PDS data, public (no licence text seen)opentruth to make
two orthos per DTM under different sun angles: a known warp tests illumination change; after HiRISE
Same-DTM ortho pairs with different sun angles: a strong illumination-change registration test.Known warp for exact truth; crater rims (Robbins DB) or Apollo hardware as point targets
Lunar cross-view geo-localisation benchmark (Nguyen, Märtens, Chin 2026)rendered 360-degree lunar surface panoramas vs overhead tiles from a high-resolution lunar terrain model10,438 panorama-tile pairs + overlapping tilesexact centre location by renderingone imagehiddenunknownunknownleft out
its images or its truth are not released
Retrieval, not pixel registration; fails the pixel criterion.Ground panoramas vs overhead tiles is retrieval, not pixel registration: skip.
Major TOM (Core S1/S2/DEM grids)global Sentinel-2 L1C/L2A, Sentinel-1 RTC and Copernicus DEM cut on a common gridmillions of co-located S1-S2-DEM cellsco-located on a shared grid (by construction); not checked per cellone imagedirectCC BY-SA 4.0 (HF card)opentruth to make
Sentinel-1 RTC and Sentinel-2 on one grid (CC BY-SA 4.0, ungated HF, 1068 px cells): a known warp gives a SAR-optical synthetic truth; co-location good to about a pixel; first SAR-optical synthetic
Ready-made multimodal same-grid pairs for building known-warp benchmarks at scale.Known warps per cell; SAR-optical check points on 50 cells
MapData (MapGlue)Electronic map to visible satellite image, 233 sites worldwide, same place121,781 pairs 512x512; test 1,910 pairs public, full set on acceptanceHomography: 200 manual points per large image, template refinement to 1 px, a-contrario check and manual recheckboth imagesdirectNot statedunknowntruth to make
map-image patches aligned from 200 manual points per scene, scored through synthetic homographies: known warp; Google Maps tiles, no licence; low
Map-to-image; Google Maps source, third-party terms.Two people click 20 check points on 50 test pairs
Mars 2020 TRN HiRISE and CTX orthomosaics and DTMs (Jezero)Jezero crater: 7 HiRISE stereo models (14 images) at 0.25 m ortho / 1 m DTM; 3 CTX stereo pairs (LVS map) at 6 mindividual HiRISE orthos (overlapping) and HiRISE vs CTX LVS mapSOCET SET bundle adjustment with manual tie points, pc_align to CTX LVS DTM, joint BA, then a 2D affine to CTX; flight-qualified for Perseverance TRNboth imagesdirectCC0 1.0 Universal (USGS data release DOI 10.5066/P9QJDP48)opentruth to make
one site, mission-controlled HiRISE/CTX mosaics: known warp; low
Mission-grade controlled multi-resolution reference with documented control chain.
MARTIAN (HiRISE-based aerial Mars rendering) and MoonAnything (rendered lunar stereo)synthetic aerial views rendered from HiRISE DTM+ortho (Mars) or from real lunar topography (Moon) under controlled illuminationrendered views with exact poses and depth vs orbital mapexact by construction (rendering)both imagesdirectMARTIAN open source (licence not checked); MoonAnything not checkedunknownwaits for a tool
rendered aerial views with exact poses against an orbital map; waits for 2D/3D pose error
Synthetic; useful for illumination-robust matching with exact truth.
Maxar Open Data ProgramWorldView/GeoEye pre- and post-event scenes for disasters, 0.3-0.5 m, many events since 2017pre/post pairs per event; multi-angle viewsnone (orthorectified with provider RPC+DEM; residual offsets of metres between dates common)one imagedirectCreative Commons Attribution Non Commercial 4.0non-commercialtruth to make
pre/post VHR scenes of many events (CC BY-NC 4.0, anonymous COG): two people placing 20 ground points per pair on 10 events give an expert multi-temporal truth; also the optical side of the SAR recipe
Non-commercial licence; the base imagery of most disaster registration sets.Two operators click 20 ground-level points on 10 pre/post pairs; DSM from Copernicus to model parallax
MESSENGER MDIS global controlled mosaics (Mercury)Mercury MDIS monochrome and colour global basemaps (~166 m/px) controlled by a global networkoverlapping MDIS frames; mosaic vs DEMglobal photogrammetric control (USGS, unverified figures)one imagedirectNASA PDS public (unverified)opentruth to make
controlled MDIS mosaics (page not confirmed): known warp; low
Primary page not confirmed in this sweep.Control-network tie points as truth
METU-VisTIR (XoFTR)Ground-level visible to thermal infrared, 6 outdoor scenes, cloudy and sunny~2.5K pairsCamera poses (relative pose truth)both imagesdirectCreative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0)non-commercialwaits for a tool
ground-level visible-thermal pairs with relative-pose truth, not remote sensing; waits for 2D/3D pose error
Fails scope: not remote sensing; listed because XoFTR is a key baseline.Not needed for remote sensing; skip
MRSIDatasets (CoFSM multimodal set, Yao et al.)Six types: multi-temporal optical, infrared-optical, depth(LiDAR)-optical, map-optical, SAR-optical, day-night; remote sensing, same scene60 pairs (10 per type) with truth; free zip has ~47 pairs, images onlyCorresponding ground-truth points per pair (in the IEEE DataPort version); how they were made not statedboth imagesaccountNot stated; truth files need an IEEE DataPort subscriptionunknownneeds an account
the download needs an account
MRSIDatasets.zip on skyearth.org has no truth files; truth sits on ieee-dataport.org/documents/cofsm.Two people click 20 check points on each of the 60 pairs and publish them openly
MU-Net multimodal registration sets (Ye Yuanxin)optical-optical (WHU 0.075 m 2012/2016), optical-infrared (Landsat-8 band 2 Jul 2020 vs band 5 Feb 2021, 30 m), optical-SAR (Sentinel-1/2 2021, 10 m), optical-map (Google, Tokyo, 1 m)~5,000 train + 500-800 test per type, 512x512precisely aligned source pairs, then 10 random affine transforms each (exact)both imagesdirectnot checkedunknowntruth to make
aligned multimodal pairs with random affines: smooth known warp; licence not checked; low
Typical synthetic-affine protocol over aligned multimodal pairs.
Multi-date WorldView-3 RPC bundle-adjustment sets (Marí et al. 2026)seasonally diverse WorldView-3 collections of 39-42 views per site (likely DFC2019/IARPA sites, unverified)all pairs within each collectionGCP-free; geometric consistency (reprojection) errors, LiDAR where availableone imagedirectsource imagery public (IARPA/DFC terms)unknownleft out
no image twice: a method paper on IARPA/DFC imagery that SatDepth already holds
Shows multi-date matching is the bottleneck for RPC refinement.Surveyed GCPs (Jacksonville set used by SatDepth) as absolute check
Multi-epoch historical aerial images: Pezenas, Frejus, Kobe (Zhang, Rupnik, Pierrot-Deseilligny)historical aerial blocks (Pezenas 1971/1981 vs IGN 2015; Frejus 1954/1966/1970 vs 2014; Awaji/Kobe 1991 vs 1995 earthquake)inter-epoch image pairs within each blockmodern epoch orientation fixed as truth; 15 manually measured check points per site (0.17 m Frejus, 0.55 m Pezenas); JMA earthquake displacement for Kobeboth imageshiddennot released ('will be soon published in an open-access repository')unknownleft out
its images or its truth are not released
Data still not released per README (Oct 2026).Ask authors; IGN historical photos (remonterletemps) can rebuild Pezenas/Frejus with IGN geodetic points
MultiEarth 2022/2023 (Amazon, Sentinel-1/2, Landsat 5/8)Amazon time series: Sentinel-1 (780,706 images, 2014-2021), Sentinel-2 (5.2 M), Landsat 5 (3.4 M, 1984-2012), Landsat 8 (2.1 M); SAR-to-EO translation and deforestation tasksSAR vs EO same-date chips (5,000 test pairs for SAR-to-EO); decades of multi-date chipsspatially and temporally aligned 256x256 chips from GEE (by construction)one imagedirect'freely available for development, research, and educational purposes' (Google Earth Engine licence agreement)non-commercialtruth to make
same-grid Amazon chips (GEE terms, non-commercial): known warp; low
Huge same-grid SAR/optical chip pool over 40 years.Known warps on chips; SAR-optical check points on 50 chips
Multimodal Image Matching Datasets (Jiang, Ma et al. 2021 review)18 multimodal pair types across medicine, remote sensing (SAR-optical, IR-optical, map-optical, day-night, multi-temporal) and computer vision164 pairs in total (remote-sensing share is a subset)15-20 manually labelled matched landmarks per pair; affine matrix for pairs that are linearboth imagesdirectnot stated (unverified)unknownleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): its licence is not shown to allow commercial use (not stated (unverified)). It was proposed as: 41 remote sensing pairs of 7 types with 15-20 manual landmarks each, direct from GitHub; linear: an affine leaves 1-3.5 px, the landmarks' own spread
Mixed domains; only the remote-sensing pairs fit this list. status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Remote sensing registration review 2026).
MultiResSARSAR (Sentinel-1, HT1-A, GF-3, Umbra 0.16 m) to Google Earth optical, 6 scene types, same scene10,850 pairs 512x512 (850 Umbra ultra-high-res); only 1,100 public so farHomography per pair: automatic registration then experts pick control points, error within 1 pxboth imagesdirectNot statedunknownneeds an account
Baidu only; 1,100 of 10,850 pairs public, delivered aligned, the expert control points not released
Baidu link only. Rest of data promised later. status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (SAR-optical review 2025 (MultiResSAR)).Ask for the expert control points as check points; add 20 independent points on 50 test pairs
MultiResSAR (Multi-Resolution-SAR dataset)SAR from four commercial satellites (incl. Umbra 0.16 m) to optical, 0.16-10 m; urban, rural, plain, hill, mountain, water10,850 pairs 512x512 (850 UHR test pairs at 16 cm); 1,100 public so farMI + RANSAC coarse, then 8-12 manual control points, 0.16 m subset re-checked (Journal of Radars 2026 table)one imagedirectnot stated (unverified)unknownleft out
duplicate of the row MultiResSAR
Only part released; Baidu only.Release the control points with the pairs; known warps
MuRA-T (SpaceNet 7 + Landsat + Sentinel aligned stacks)PlanetScope (4 m) vs Landsat (30 m) vs Sentinel-2 (10 m) time series over 60 SpaceNet 7 AOIs60 AOI datacubes, monthlycross-satellite alignment by MuRA framework with AROP-style polynomial misalignment modelsone imagedirectunknownunknownleft out
its images or its truth are not released: download location not found
Download location not confirmed (unverified).Check alignment with SpaceNet 7 building centroids visible at 10 m; known warps
Murray Lab global CTX mosaic of Mars (V01)86,571 CTX images (2006-2020) blended into a 5 m/px global mosaic, 88S-88Noverlapping source images (all registered to each other); mosaic vs THEMIS/MOLA/HRSC/HiRISE3,147,169 image-to-image tie points with residuals shipped as shapefiles per tile; registered to controlled basemaps, absolute error < 200 mboth imagesdirectno licence text seen; acknowledge Caltech, MSSS, JPL; PDS archive DOI 10.17189/m23v-ae12 (NASA data, generally public)opentruth to make
3.1 million tie points, 99% automatic, for a translation-only co-registration (Dickson et al. 2024): people would check them; 5.6 TB; low
The largest open co-registration product with released tie points: real out-of-core scale and a ready-made truth set.Use the shipped tie points as truth for re-registering raw CTX pairs; hold out 10% of points as check
MUUFL Gulfport hyperspectral and LiDARairborne CASI hyperspectral (72 bands) and LiDAR over University of Southern Mississippi Gulf Park campus, multiple flightsHSI-LiDAR pairs per campus flight; several HSI flights of the same campusco-registered rasters by provider; scene labels; GPS-surveyed target locations (KML)one imagedirectMIT License (repository)opentruth to make
small campus HSI and LiDAR rasters with GPS-surveyed targets: target points; low
Overlapping flights give multi-temporal HSI; known co-registration imperfections reported by users (unverified).Surveyed target positions as point truth in both HSI and LiDAR; known warps
NAIP (USDA aerial orthoimagery)leaf-on aerial RGB+NIR orthophotos, 0.3-1 m, continental US, repeat every 2-3 years per statesame place in successive NAIP years; NAIP vs Sentinel-2/Umbra/Maxarorthorectified to USDA accuracy spec (e.g. within ~6 m of reference, unverified); no per-pair truthone imagedirectPublic Domain with Attributionopentruth to make
public-domain orthos every 2-3 years: people points between years; also the open optical side for US Umbra/Capella sites
Multi-year aerial pairs with seam/mosaic offsets: realistic non-rigid misregistration.Known warps; road-junction check points vs USGS 3DEP-derived features
NBU_PansharpRSData (Meng et al. 2020)MS vs PAN from six satellites (IKONOS, QuickBird, GaoFen-1/2, WorldView-2/3; per paper)about 2,270 MS/PAN pairs (per paper, unverified)same-platform co-registrationone imagedirectacademic purpose only, cite papersnon-commercialneeds an account
Baidu only; same-platform MS/PAN
Baidu only.Known warp between MS and PAN
Night-light imagery registration (Luojia-1, SDGSAT-1 GLI, VIIRS DNB): no public benchmarknight-time light images to each other (mosaics) and to daytime optical or OSM roads; 10-750 m GSDnone shipped; Luojia-1 test used 275 scenes over China (Guan 2022)none public; papers use automatic tie points or OSM-road-derived GCPsaccountVIIRS DNB NASA/NOAA open; Luojia-1 free download from Hubei centre; SDGSAT-1 open via CASEarth (registration needed) (unverified)unknownneeds an account
the download needs an account
Listed as a gap: no shipped pairs with truth were found.Pair SDGSAT-1 GLI night scenes with same-day daytime Sentinel-2; two operators click 20 street-light/road-junction points on 10 pairs; or use OSM junctions as GCPs
NIR-VIS-RS (RIZER dataset)Infrared to visible remote sensing (desert, coast, city, mountain, aerial)Not stated on pages seenReal transformation per pair (method unverified)requestNot statedunknownneeds a request
the data needs a request or a data use agreement
Repo links only model weights; dataset download not seen (unverified).Two people click 20 points on each test pair
OpenAerialMapcommunity-uploaded drone and aerial orthomosaics worldwide (cm-level), often several flights of the same placemulti-flight pairs over the same site; drone vs satellite basemapnone (uploader georeferencing)one imagedirectCC BY 4.0 for imagery uploaded to OAM (unverified)opentruth to make
community orthomosaics, several flights of a place: people points; licence per upload unverified; low
Primary page not confirmed in this sweep.Two operators click 20 points per pair on 10 overlapping mosaics
OpenEarthMap-SAR (DFC 2025 track 1)VHR SAR (Umbra) and aerial/satellite optical 0.15-0.5 m, 35 regions in Japan, France, USA5,033 images 1024x1024 (SAR-optical aligned tiles)co-registered SAR-optical tiles (method in paper, unverified); land-cover labels (partly manual, rest pseudo)one imagedirectZenodo record licence not checked (Zenodo blocked this sweep)unknowntruth to make
pre-aligned VHR SAR-optical tiles (little room): known warp; Zenodo licence not checked; low
Companion of BRIGHT (track 2).Check points on 30 tiles by two operators; known warps
OPERA CSLC-S1 (coregistered geocoded Sentinel-1 bursts)Sentinel-1 IW SLC bursts geocoded onto fixed UTM grids so all dates are stackedburst time series (co-registered by construction for InSAR)geocoding with precise orbits and DEM; geolocation validated with corner reflectors (product docs, unverified)one imageaccountCreative Commons BY 4.0 (OPERA products, per AWS registry for RTC; CSLC assumed same, unverified)openneeds an account
the download needs an account
Ready SAR stacks where co-registration is already done: reference for testing SAR-to-SAR registration of raw SLCs.Raw SLC vs CSLC grid as reference transform; corner reflectors (Queensland array) as point truth
OPERA RTC-S1 (NASA, Sentinel-1 terrain-corrected bursts)Sentinel-1 IW bursts, radiometrically terrain corrected, on fixed UTM 30 m grid, near-globalsame burst on many dates (co-registered by construction on the grid); RTC vs Sentinel-2/HLSgeocoded with DEM to a fixed grid; validated geolocation (product specs, unverified)one imageaccountCreative Commons BY 4.0openneeds an account
the download needs an account
Same-grid SAR stacks: SAR-to-SAR and SAR-to-optical truth by construction.Known warp between dates or vs HLS; corner reflector sites as point truth
OrthoLoC (UAV 6-DoF localisation against orthophotos and DSMs)UAV photos (2D) against 5 cm orthophotos and digital surface models (2.5D/3D) from public geodata, Germany and USabout 16,400 images in 47-51 regions of 19 cities; train 13,347 / val 1,489 / out-of-place test 1,5916-DoF poses from georeferenced multi-view photogrammetry with RTK or manually annotated ground control pointsboth imagesdirectCC BY-NC-SA 4.0 (repository badge); data on TUM websharenon-commercialwaits for a tool
UAV photos against orthophoto and DSM: pose truth; CC BY-NC-SA; waits for 2D/3D pose error
real; geodata and query images differ in date and sensor, a true cross-source 2D/3D test
OS-Eval (Xiang et al. 2023)Sub-meter optical (WorldView-2/3, SuperView-1, Google Earth) to GaoFen-3 SAR, airports, residential, industrial, same sceneHundreds of pairs (review says 1,232); relative and absolute subsets with RPCs, DEM, DSMPoles with clear structure in both optical and SAR (README: "meta poles") used as ground-truth pointsboth imagesdirectThe contents of this dataset are subject to CC BY-NC-ND (Creative Commons Attribution-NoDerivatives 4.0 International).non-commercialtruth to make
two sites only (11 optical x 13 SAR, 19 SAR vs Google), a few light-pole points per image; people would add points; CC BY-NC-ND; low
Rare: physical markers in both modalities. Licence text mixes NC-ND and ND wording. status change: Round 4 count: listed in 2 of the 38 round-4 surveys and benchmark papers (Remote sensing registration review 2026; SAR-optical review 2025 (MultiResSAR)).
OSCD (Onera Satellite Change Detection)Sentinel-2 13-band pairs (10/20/60 m), 24 cities worldwide, 2015-201824 pairs (14 train, 10 test), ~600x600 px each'registered pairs' by provider (Sentinel-2 L1C geolocation, no residual given); pixel change labels (test labels now open)one imagedirectnot stated on page; Copernicus Sentinel data open (dataset licence on IEEE DataPort, unverified)unknowntruth to make
24 Sentinel-2 pairs, pre-GRI geolocation: known warp; low
Early Sentinel-2 (pre-GRI) products can be off by a pixel or more between dates.Re-derive pairs from GRI-refined L1C (baseline 04.00+) and measure residual with AROSICS; known warps for exact tests
OSdataset (Xiang et al. 2020)GaoFen-3 spotlight SAR (1 m) to Google Earth optical (1 m), 20 scenes in China, France, USA, India, same scene2,673 pairs 512x512 or 10,692 pairs 256x256Pairs co-registered by the authors' phase-congruency method, then visually checked; no independent pointsone imagedirectOur dataset is shared under the open-access license CC-BY-NC.non-commercialtruth to make
GF-3/Google Earth patches aligned by the authors' matcher: known warp plus a check of the base; CC BY-NC, Google Earth terms; low
Truth is algorithmic co-registration; most papers add synthetic rotations or flows. Google Earth imagery terms also apply. status change: Round 4 count: listed in 2 of the 38 round-4 surveys and benchmark papers (Remote sensing registration review 2026; SAR-optical review 2025 (MultiResSAR)).Apply known affine plus smooth deformation to test pairs (as GDROS does); verify base alignment with 20 expert check points on 50 pairs
OSDataset2.0 (Xiang et al., Journal of Radars 2026)SAR to optical, patch level (14 countries) plus one full scene pair (0.43 m)6,476 registered 512x512 patch pairs; 1 scene-level pairScene pair: high-precision correspondences (72 sub-pixel points per the user manual); patch-level truth method not statedboth imagesdirectNot stated on the page seen (unverified)unknowntruth to make
one full scene pair (35k px at 0.43 m) with 72 pole points plus patches aligned by matching: one scene cannot make ten pairs; 8.86 GB zip on radars.ac.cn (login not checked); low
New 2026; download page radars.ac.cn/web/data not visited. Includes the original OSdataset.Two people click 20 check points on 20 patch pairs to measure patch-level alignment error
OSM building label-noise dataset (DLR, Zenodo 4446737)rendered OpenStreetMap imagery vs building footprint labels with controlled noise types, incl. shiftstiles with clean and noisy label sets (count unverified)exact: labels and imagery from the same OSM source, then known shifts and other noise addedboth imagesdirectnot checked (Zenodo blocked during this sweep; unverified)unknownleft out
no second image: rendered map imagery against its own labels
Imagery is rendered map, not real sensor data; weak for registration but exact by construction.
PanCollection (WV2, WV3, QuickBird, GaoFen-2 pansharpening)multispectral (4-8 bands) vs panchromatic of the same satellite, 4 sensorstraining sets 5-6 GB per sensor; 20 reduced-res and 20 full-res test examples per sensorMS and PAN from the same platform, co-registered by provider; reduced-res sets made by Wald protocolone imagedirectnot stated (imagery from Maxar/GaoFen; unverified)unknowntruth to make
same-platform MS and PAN: a known warp across resolutions; licence not stated; low
Small residual MS-PAN misregistration exists in real products; the set assumes alignment.Known warp of MS against PAN gives exact cross-resolution multispectral truth
Planet data (PlanetScope/SkySat; NICFI tropical mosaics; Education and Research programme)PlanetScope 3-4 m daily scenes and monthly basemaps, SkySat 0.5 mdaily multi-date pairs; PlanetScope vs Sentinel-2none per pair (Planet orthorectification; known inter-scene misregistration of a few metres)one imageaccountPlanet licence terms (NICFI: participant licence, non-commercial; E&R: research only) (unverified wording)non-commercialneeds an account
the download needs an account
Commercial imagery; open copies exist only inside SpaceNet 7/DynamicEarthNet.Sentinel-2 GRI-refined scenes as reference; manual check points
Queensland corner reflector array (Geoscience Australia, Surat Basin)40 surveyed trihedral corner reflectors over ~130 x 130 km, visible in Sentinel-1, TerraSAR-X and other SAR imagesany SAR acquisition over the array (repeat passes since 2014) vs surveyed truthgeodetic survey of reflector apex points (GNSS + total station, 2018; plate-motion velocities); sub-cm to cm levelboth imagesdirectGeoscience Australia record (licence text not seen; GA data usually CC BY 4.0, unverified)unknowntruth to make
40 surveyed corner reflectors seen in any Sentinel-1 pass: reflector peaks on both images make point pairs; one site; low
Physical markers: an exact point truth for SAR geolocation and SAR-to-SAR co-registration.
QXS-SAROPTGaoFen-3 spotlight SAR (1 m) to Google Earth optical, San Diego, Shanghai, Qingdao, same scene20,000 patch pairs 256x256Manually located matching points on sub-regions, then registration by software; patches co-registeredone imagerequestThe QXS-SAROPT dataset under open access license CCBY (paper)openneeds a request
the data needs a request or a data use agreement
Download link sits behind a short survey (wenjuan.com). Optical from Google Earth: third-party terms. status change: Round 4 count: listed in 2 of the 38 round-4 surveys and benchmark papers (Remote sensing registration review 2026; SAR-optical review 2025 (MultiResSAR)).Apply known rotations and smooth warps; two people click 10 check points on 50 pairs to measure base error
Real HSI-MSI-PAN fusion dataset (Li 2024)real hyperspectral, multispectral and panchromatic satellite images of the same area (sensors per paper, unverified)HSI-MSI, HSI-PAN, MSI-PAN pairs (count unverified)spatially registered by the author (processing code released, github rs-lsl/CSSNet)one imageaccountunknownunknownneeds an account
the download needs an account
Real, not simulated, cross-sensor spectral pairs; alignment quality not quantified.Known warp between HSI and MSI for exact truth; manual check points on 10 scenes for real residuals
ReBO (Repairing Buildings in OSM, DragOSM)OpenStreetMap building polygons vs off-nadir aerial/satellite images, 41 cities5,473 images, 179,265 buildingsmanually corrected roof and footprint polygons next to the original OSM polygon (per-instance offset)both imagesdirectApache-2.0 (HF card kevinlikai/ReBO)openleft out
no second image: OSM polygons corrected against an image
Vector-to-image alignment with instance-level manual truth; offsets come from parallax and map error, not image-to-image warps.
REMM rotation-scale benchmarkOptical-SAR (SEN1-2, GF-3 OS) and optical-NIR (VEDAI aerial, RGB-NIR scenes)316 base pairs expanded to 33,180 test pairsKnown homography: 36 rotations x 3 scale ranges applied to aligned pairsboth imagesdirectNot statedunknownneeds an account
Baidu only; synthetic rotations and scales on pre-aligned pairs
Baidu link only; synthetic geometry on pre-aligned pairs.Add non-rigid warps; check base alignment of the 316 pairs with points
RIFT and LNIFT test sets (Li et al.)SAR-optical, infrared-optical, depth-optical (LNIFT); six multimodal types (RIFT)LNIFT: 4,000 pairs rebuilt from QXS-SAROPT, OSdataset, DIML RGB-D, LLVIP; repos hold only demo pairsIdentity (pre-aligned) or known random rotation applied to aligned pairsboth imagesdirectNot statedunknownleft out
its images or its truth are not released: demo pairs only, the 4,000 pairs to be rebuilt
Not released as a dataset; rebuild from sources. Depth and LLVIP parts are not remote sensing.Rebuild with fixed seeds and add smooth warps; publish the transforms
RIFT multimodal remote sensing pairs (Li, Hu, Ai 2020)optical-optical, infrared-optical, SAR-optical, depth-optical, map-optical, day-night60 pairs (10 per type)affine from 5 evenly spread sub-pixel correspondences per pair, picked by handboth imagesdirectnot stated on the README (unverified)unknowntruth to make
60 pairs with 5 manual points each: people would add 10-15 per pair; after the Jiang/Ma set, which covers the same types
Predecessor of the SRIF set; 5 points per pair is thin truth.Add 15-20 points per pair by two people; report affine residuals
Robbins & Hynek Mars crater database (2012)~384,000 Martian craters >= 1 km with rim-fit centres and diameterslandmark list usable on any Mars mosaic pairmanually traced rims on THEMIS daytime IR mosaic (control frame of THEMIS)directpublic (USGS Astropedia / PDS; unverified)openleft out
no second image: a crater list in one frame
Primary page not confirmed in this sweep; landmarks are in one frame, not paired.Crater centres re-fitted on both images of a pair give natural point landmarks; ellipse fits give sub-pixel centres
Robbins lunar crater database (2019)~1.3 million lunar craters complete to ~1-2 km, rims traced on WAC, Kaguya TC and LOLA/TC DTM hillshadeslandmark list for any lunar image pairmanual rim tracing (~2.5 px per vertex), circle and ellipse fits; LOLA control framedirectpublic (USGS Astropedia / PDS4; no licence text seen)openleft out
no second image: a crater list in one frame
Accurate in aggregate, not per crater (database text).Re-fit crater rims in both images to obtain paired centres as landmarks
S2Lookingside-looking (off-nadir, mean |9.86| deg) GaoFen/SuperView/BeiJing-2 bitemporal pairs, 0.5-0.8 m, rural areas worldwide, 1-3 years apart5,000 pairs 1024x1024pairs registered by provider (method in paper, unverified); building change labelsone imagedirectnot stated on README (unverified)unknowntruth to make
off-nadir bitemporal pairs registered by the provider: people points on ground features; licence not stated; low
Off-nadir parallax makes pixel-perfect alignment impossible on buildings: a natural non-rigid case.Unchanged ground-level points (road crossings) clicked by two operators on 30 pairs; DSM unavailable
SARLO-80 (Umbra SLC slant-range + aligned optical + captions)Umbra SLC SAR in slant-range (80 cm) to Google satellite optical, 257 locations in 72 countries119,566 triplets 1024x1024 (SAR, optical metadata, captions)None: optical warped by a corner affine from SICD geolocation onlyone imagedirectSAR: Umbra SICD under CC-BY-4.0, redistributed with attribution; optical not redistributed (Google TMS)openleft out
no second image delivered: the optical side must be rebuilt from Google tiles
Fails truth; optical must be rebuilt from Google tiles. Slant-range geometry is unusual.Apply known warps to rebuilt pairs; or two people click 20 points on 20 pairs
SARpticalTerraSAR-X spotlight SAR (~1 m) to UltraCAM aerial optical (0.2 m), dense urban Berlin, same scene~10,108 patch pairs 112x112Correspondence from co-registered TomoSAR and optical 3D point clouds (Wang and Zhu 2018)both imagesdirectNot statedunknowntruth to make
112 px SAR-optical patches matched through TomoSAR and optical point clouds: too small for a field, larger crops need the source scenes; licence not shown; low
Small patches, heavy overlap; good for SAR layover cases. Licence not shown. status change: Round 4 count: listed in 2 of the 38 round-4 surveys and benchmark papers (Remote sensing registration review 2026; SAR-optical review 2025 (MultiResSAR)).Use the 3D point correspondences as dense truth; add known offsets for registration tests
SatDepthmulti-date, multi-view WorldView-2/3 panchromatic images (0.25-0.5 m) over 4 AOIs (Jacksonville, Omaha, UCSD, San Fernando), large view and track angle differencesmany thousand image pairs from 198 PAN images (tiles); Jacksonville train, others testdense pixel correspondences from bundle-adjusted RPCs + stereo-fused DSM ('SatDepth maps'); checked with 76 surveyed GCPs (~2000 image annotations): relative 2D error 2.22 px meanboth imagesdirect: https://engineering.purdue.edu/RVL/Database/satdepth/dataset/ (tar parts, ranges accepted); index and GCP archives on Google Drivenone stated for the data; code Apache-2.0; imagery courtesy DigitalGlobe through IARPA's public releasesaskleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): its licence is not shown to allow commercial use (none stated for the data; code Apache-2.0; imagery courtesy DigitalGlobe through IARPA's public releases). It was proposed as: The dataset is 611 GB in 123 parts of one plain tar (5.24 GB each) on Purdue's server, which accepts byte ranges, so members can be reached by walking the tar headers, but the server drops connections after a few dozen requests. The dense correspondences come from bundle-adjusted RPCs and a stereo DSM (a program); people annotated GCPs only for Jacksonville (gcps.tar.gz on Google Drive). No data licence on the download page or the repository (Apache-2.0 covers the code); the imagery is CORE3D and IARPA MVS3DM WorldView.
Rare dense truth for real non-orthorectified satellite pairs; truth limited by DSM accuracy (MAE 0.59 m vs LiDAR).
Satellogic EarthViewSatellogic ~1 m RGB+NIR images with matching Sentinel-1/2 and NEON-style data (per card, unverified), multi-dateco-located Satellogic vs Sentinel stacksco-located by georeference; no check pointsone imageaccountCC BY 4.0 (gated: auto approval on Hugging Face)openneeds an account
the download needs an account
Gated on HF (click-through), so 'account'.Known warps; check points vs Sentinel-2 GRI chips
SCDF co-registration test pairs (Sentinel-2, Landsat-8/9, NAIP)large optical scenes up to 8192x8192: Sentinel-2, Landsat 8/9, NAIP, cross-sensor and multi-date584 pairs with constructed ground truth (method in paper)constructed ground truth (details not seen)hiddenpaper CC BY 4.0; data release not linkedunknownleft out
its images or its truth are not released
Directly about large-scene dense co-registration; data not released.Ask authors for the pairs; else rebuild from open COGs with known warps
SECOND (semantic change detection)aerial bitemporal pairs, Chinese cities, 0.5-3 m4,662 pairs 512x512co-registered by provider; semantic change labels on both datesone imagedirectunverifiedunknowntruth to make
aerial change pairs aligned by the provider (page not confirmed): known warp; low
Page not confirmed in this sweep.Known warps
SEN1-2 (and SEN12MS, SEN12MS-CR)Sentinel-1 SAR (VV) to Sentinel-2 RGB, global, all seasons, same scene; SEN12MS adds 13-band S2 and MODIS land coverSEN1-2 282,384 patch pairs 256x256; SEN12MS 180,662 tripletsCo-registered by geocoding (terrain-corrected S1, orthorectified S2); no pointsone imagedirectRights: by, http://creativecommons.org/licenses/by/4.0opentruth to make
the same S1/S2 pairing as Major TOM in 256 px patches (CC BY 4.0): known warp; after Major TOM
Fails truth on its own: residual geocoding error is near the 10 m pixel. SEN12MS-CR (cloud removal) is the same family. status change: Round 4 count: listed in 2 of the 38 round-4 surveys and benchmark papers (Remote sensing registration review 2026; SAR-optical review 2025 (MultiResSAR)).Apply a known affine plus smooth warp to test patches; measure base misalignment on 20 patches with expert points
SEN12MS-CR and SEN12MS-CR-TSSentinel-1 SAR + cloud-free and cloudy Sentinel-2, 10 m; -TS adds 30-date time series per area122,218 triplets 256x256 (CR); TS: 53 regions x 30 dates (unverified count)co-registered by construction on the Sentinel grid; no check pointsone imagedirectCC BY 4.0opentruth to make
cloudy and clear Sentinel-2 with Sentinel-1 on one grid: known warp tests occlusion; after Major TOM
Made for cloud removal; useful as multi-date same-grid stacks for registration-by-construction tests, including cloud occlusion.Known warp between dates or between SAR and optical; cloudy vs clear image pairs test robustness to occlusion
Sentinel-1 GRD COGs on AWS (sentinel-s1-l1c)global Sentinel-1 GRD archive (IW/EW, VV/VH) converted to COG; same orbit track repeats every 6-12 daysunlimited multi-date SAR pairs; SAR vs Sentinel-2 opticalnone (GRD is in slant-range-projected ground geometry with GCP grid, not terrain-corrected)one imagedirect'Access to Sentinel data is free, full and open ...' (Copernicus legal notice)opentruth to make
GRD in its own geometry: terrain correction then known warp; Major TOM already gives RTC; low
Raw-ish geometry: registration to optical needs terrain correction, a realistic hard case.Terrain-correct with Copernicus DEM, pair with Sentinel-2 on the same grid; known warps; corner reflectors where present as truth
Sentinel-2 Global Reference Image (GRI) and GRI GCP databaseSentinel-2 B04 reference layer (2015-2018) and GCP chip database for registering any Sentinel-2 date (multi-temporal)Global tiles; up to 400 GCPs per square degree (57x57 chips)Geometric refinement by spatio-triangulation on reference GCPs; absolute 7 m CE95, internal 1 m CE95both imagesaccountThe Multi-Layer L1C GRI can be downloaded and publicly distributed to the international user community.openneeds an account
the download needs an account
Not pairs as such; a reference to build multi-temporal pairs with check points. CDSE S3 needs credentials. Points are image-derived, not surveyed.Pick test tiles over areas with surveyed GCPs (national geodetic networks) for independent checks
Sentinel-2 L2A COGs on AWS (Earth Search, Element 84)full Sentinel-2 MGRS tiles, 13 bands at 10/20/60 m, global, 2015-now; same tile on many datesunlimited: any two dates of a tile, or band-to-bandgeometric refinement against the GRI since Aug 2021: absolute < 6 m, multi-temporal < 5 m (95%), multispectral < 0.3 px (ESA quality report); no per-pair truthone imagedirect'Access to Sentinel data is free, full and open ...' (Copernicus Sentinel Data Legal Notice)opentruth to make
two GRI-refined dates of one tile (multi-temporal under 5 m at 10 m): a known warp by blocks on a 120 Mpx tile; second synthetic
Best open source of large real scenes; pre-2021 (unrefined) acquisitions have larger multi-date offsets, a useful hard case.Two refined L1C/L2A dates of one tile = residual < 5 m truth; apply a known smooth warp to a full tile chunk by chunk for exact truth; GRI GCP chips as independent check
SkyReg (pixel-wise drone-satellite geo-registration)Drone views (rendered in Google Earth Studio) to orthorectified Bing or perspective satellite views; Chicago, Seattle, San Francisco, landmarksTrain ~115k drone images; bench 13,191 (Chicago) + 950 (suburban)Per-pixel lat/lon from USGS LiDAR (urban) or SfM with fixed cameras (landmarks, suburban)both imagesdirectNot statedunknownwaits for a tool
rendered drone views over 3D cities with per-pixel lat/lon from LiDAR or SfM: perspective views with occlusions; licence not stated; waits for 2D/3D pose error
Drone images are renders, not real flights. Dense truth, large viewpoint change.Real drone flights with RTK GCP targets on a subset
SOMA-1MSAR (Sentinel-1, PIESAT-1, Capella, 0.5-10 m) to Google Earth optical, 12 scene types1,300,954 pairs 512x512; test set released, rest progressiveAutomatic coarse-to-fine matching framework, quality cleaning; no independent points statedone imagedirectSOMA-1M dataset resources will be released progressively through this repository in accordance with applicable third-party data licenses and copyright requirements.unknowntruth to make
SAR-optical pairs aligned by an automatic matcher, release progressive under third-party terms: known warp; low
Truth is algorithmic; mostly a training set.Experts click 20 check points on 100 test pairs; apply known warps for scoring
SomBench (co-registered lunar multi-instrument tiles)30+ co-registered lunar layers from ten instruments (LRO, Kaguya, GRAIL, Lunar Prospector), 1 m to 20 km/px, 82 deg latitudemultimodal tile bundles (~2 million per companion paper)co-registered by construction on Lunar Transverse Mercator zonesone imagehiddenunknownunknownleft out
its images or its truth are not released
Release status not checked; registration quality between instruments not stated in abstract.Known warps across layers for exact multimodal lunar truth
SOPatchSAR (Sentinel-1, GaoFen-3) to optical (Sentinel-2, Google Earth), from WHU-SEN-City, OSdataset, SEN1-2>650,000 patch pairs 64x64Source pairs re-registered precisely by the authors, then known position and rotation noise added (HPatches style)both imagesdirectNot statedunknownneeds an account
Baidu only; 64 px descriptor patches
Patches too small for whole-image registration; descriptor benchmark. WHU-SEN-City (34 Chinese cities, S1/S2) is one source, no licence either. status change: Round 4 count: listed in 2 of the 38 round-4 surveys and benchmark papers (Remote sensing registration review 2026; SAR-optical review 2025 (MultiResSAR)).Draw larger test crops from the source scenes and apply known smooth warps
SpaceNet 4 Off-Nadir (Atlanta)27 WorldView-2 8-band images of the same Atlanta area taken the same day (22 Dec 2009) at 7 to 54 deg off-nadirany two of 27 views (351 pairs), 665 km2, tiled 450 morthorectified by provider; ~120,000 building footprints (ground footprint, same for all views)one imagedirectSpaceNet open licence (CC BY-SA 4.0 per SpaceNet, not seen on page; unverified)opentruth to make
27 same-day off-nadir views: people points on ground features (building lean is not a smooth field); low
Same-day multi-angle views: misregistration is pure parallax (building lean), a good test of non-rigid vs rigid models.Building roof corners clicked in nadir and off-nadir view by two operators on 20 tiles; or LiDAR/DSM-driven parallax field
SpaceNet 6 / MSAW (Multi-Sensor All-Weather Mapping)Capella SAR (0.5 m, quad-pol) and Maxar WorldView-2 optical over Rotterdam, same scene202 SAR strips plus ~92 km2 WV-2 in the expanded release; tiles for trainingNone for registration: products are georeferenced only; building footprints (~48,000) on bothone imagedirectThe SpaceNet Dataset by SpaceNet Partners is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.opentruth to make
Capella SAR and WV-2 over one city, georeferenced only: people tie points on tile pairs; one city; low
Fails truth; building footprints give a label check on both images.Two people click 20 ground control points (road crossings) on 10 SAR-optical tile pairs; or apply a known smooth warp to a georeferenced pair
SpaceNet 7 MUDS (Multi-temporal Urban Development)monthly PlanetScope mosaics (4 m) over 101 AOIs, ~24 months each (2017-2020)2,389 images (1024x1024); any two months per AOImosaics georeferenced by Planet (residual misregistration of a pixel or more, unverified); 11 million building footprints per month with persistent IDsboth imagesdirectSpaceNet open licence (CC BY-SA 4.0 per SpaceNet, unverified)opentruth to make
monthly 4 m mosaics co-registered to about a pixel, so building-ID centroids leave no room: known warp with the centroids moved through it; low
Persistent building IDs across months give point correspondences (polygon centroids) on every date.Centroids of the same building ID across months as dense point pairs; known warps
SRIF multimodal matching dataset (Li, Hu, Zhang 2023)Six categories: optical-optical, infrared-optical, SAR-optical, depth-optical, map-optical, day-night1,200 pairs (200 per category) 256x256Similarity transform (scale, rotation, shift) per pair: aligned pairs with a known synthetic transformboth imagesdirectNot statedunknowntruth to make
1,200 pre-aligned pairs with known similarities: check the base, add a smooth warp; after the Jiang/Ma set; low
Synthetic geometric truth on pre-aligned real pairs; base alignment not checked.Apply smooth non-rigid warps too; verify base alignment with 10 points on 20 pairs per category
SSL4EO-S12Sentinel-1 GRD + Sentinel-2 L1C/L2A, 4 seasonal dates per location, 250k locationsS1-S2 and seasonal pairs per location (264x264 px)co-located by construction (GEE export on one grid)one imagedirectCC BY 4.0opentruth to make
four seasons per location on one grid (CC BY 4.0): seasonal known warp on 264 px patches; after Sentinel-2 tiles
Seasonal 4-date stacks are good for temporal-appearance robustness with exact known warps.Known warps between seasons or modalities
Street-view to satellite sets (VIGOR, KITTI-satellite, CVUSA, CVACT): fail pixel criterionground panoramas/car camera vs aerial/satellite tilestens of thousandsGPS position (and heading) of the ground camera in the aerial tile; no pixel correspondencesone imagedirectvaries per set (unverified)unknownwaits for a tool
ground panoramas against aerial tiles: GPS and heading only; waits for 2D/3D pose error
Listed only to record that they fail the pixel-level correspondence rule.None credible beyond GPS; skip
swisstopo SWISSIMAGE 10 cm (and historical SWISSIMAGE / aerial archive)national RGB orthophoto of Switzerland: 0.1 m in plains and main valleys, 0.25 m over the Alps (2017+); older 0.25-0.5 m (1998-2016); 3-year update cyclesame place across update cycles; strong relief (Alps)positional accuracy +-0.15 m at 0.1 m GSD (2017+); +-0.25 m, or +-3-5 m in hilly terrain (2005-2016); +-0.5 m (1998-2005); no per-pair truthone imagedirectswisstopo terms of use for free geodata (page links to them; exact wording not read); Planet credit for RapidEye-sourced lake areasopenneeds a request
only the 2017+ epoch downloads directly; older epochs need a paid request, so no pair without one
Stated accuracy per epoch makes it a calibrated reference; pre-2017 epochs need a paid request.swisstopo fixed points (geodetic markers) as check points; known warps on full tiles
SYSU-CDaerial 0.5 m bitemporal pairs, Hong Kong 2007 vs 201420,000 pairs 256x256co-registered by provider; change labelsone imagedirectunverifiedunknowntruth to make
aerial change patches aligned by the provider (page not confirmed): known warp; low
Page not confirmed in this sweep. URL is the dataset's GitHub (homepage links from there).Known warps
THEMIS controlled IR mosaics (USGS)Mars Odyssey THEMIS daytime/nighttime IR at ~100 m/px, tiled by Mars quadrangleTHEMIS IR vs CTX/HRSC visible (thermal-to-visible multimodal); day vs night IRabsolutely controlled to improved Viking MDIM network; image-to-image precision sub-pixel (< 100 m)one imagedirectCC0 1.0opentruth to make
controlled THEMIS IR against controlled CTX: thermal-visible known warp; low
Thermal vs visible on Mars: a planetary analogue of optical-thermal registration.CTX mosaic vs THEMIS tile: both controlled; MOLA-derived shaded relief as third modality
Thermal and visible aerial imagery (Garcia-Moreno et al. 2020)UAV thermal (Zenmuse XT) and visible (Zenmuse X3) images of a flat campus scene to a georeferenced orthoimage80 thermal + 30 visible images, each with a homography to the orthoimageHomographies from a few manually selected correspondencesboth imagesdirectCC BY 4.0opentruth to make
one planar campus: 110 UAV images with homographies to the orthoimage from a few manual points, points not shipped: people would place 10+ per image; low
Small, planar scene; few points per homography.Click 20 points per image to refit and report residuals
UAV-VisLocnadir UAV images (0.1-0.2 m) to satellite orthomaps, 11 sites in China, various terrains6,742 UAV images vs 11 satellite mapsGPS lat/lon of image centre, height, heading; satellite map boundsone imagedirectnot stated (unverified)unknowntruth to make
nadir UAV frames with centre GPS against Google Maps tiles: people points; Google terms; low
Truth is one point per image (centre), not dense.Manual 10-point correspondences on 100 UAV-map pairs by two operators; homography from them
UAVD4L (UAV 6-DoF localisation against textured mesh and DSM)UAV query photos (2D) against a photogrammetric textured mesh and DSM (3D) with rendered database imagesreal query sets plus synthetic rendered database (2.5 km2 area)query poses from RTK-GNSS/IMU with refinement against the mesh (paper); database poses by constructionboth imagesdirectnone stated (Baidu Netdisk download)unknownwaits for a tool
UAV photos against a textured mesh and DSM: pose truth; waits for 2D/3D pose error
real queries; pose truth partly derived from the same mesh
UAVLoc-M3oblique/nadir UAV photos (Sony cameras, 300-800 m altitude) to Bing satellite maps (0.5-0.97 m), 3 regions in China, spring/summer, fog4,946 UAV images vs 13 large satellite mapsUAV GPS position and attitude (pitch, yaw, roll) from flight log; map extents; DEM tiles includedone imagedirectnot stated (ScienceDB DOI 10.57760/sciencedb.29772)unknowntruth to make
UAV photos with pose against Bing maps: people points; Bing terms; low
Truth is pose/position, not pixel correspondence.Project UAV footprint with pose + DEM, then hand-refine 10 points per image on 50 images
Umbra SAR Open DataX-band spotlight SAR, 16 cm to 1 m, >20 time-series sites + 250+ other imagesmulti-date SAR stacks per site; SAR vs open optical (NAIP, Maxar ODP)none (GEC geocoded and SICD products)one imagedirect'All data is provided with a Creative Commons License (CC by 4.0)'opentruth to make
open VHR SAR (CC BY 4.0, anonymous, 20+ time-series sites): with NAIP or Maxar ODP optical, two people placing 20+ tie points per pair on 10 sites make an expert optical-SAR truth; first people-made recipe
Time-series sites give same-geometry repeat passes, near-exact SAR-to-SAR pairs.Repeat-pass same geometry = aligned up to orbit error; known warps; check points vs NAIP
US3D / DFC2019Multi-date WorldView-3 (26 Jacksonville, 43 Omaha; Atlanta added) with airborne LiDARMany tiles x dozens of dates; epipolar pairs with disparityRPCs adjusted to LiDAR by mutual information; per-pixel XYZ and disparity from LiDARboth imagesaccountNo overall licence stated; Atlanta part: Creative Commons Attribution-ShareAlike 4.0 International Licenseunknownneeds an account
the download needs an account
Gives dense cross-date correspondences via XYZ; LiDAR not simultaneous, MI alignment automatic.Check RPC alignment with 20 surveyed or expert points per city
USGS 3DEP LiDAR (EPT/COPC on AWS)airborne LiDAR point clouds over the US; derived DSM/intensity vs NAIP or satellite imagesLiDAR-derived rasters vs optical/SAR for the same areaLiDAR is surveyed-grade georeferenced (project-level accuracy reports); no image pairs shippedone imagedirectUS Government Public Domainopenleft out
no second image: LiDAR point clouds, a tool for making dense truth
Gives an independent 3D reference to make dense truth for satellite/aerial pairs (SatDepth-style).Render LiDAR DSM/intensity; project through image RPCs to create dense correspondences for any two images
USGS Landsat Collection 2 (AWS, Planetary Computer)Landsat 4-9 Level-1/2 scenes, 30 m (15 m pan), 1982-now, WRS-2 path/row repeatsdecades-long multi-date pairs; MS vs pan; Landsat vs Sentinel-2Level-1 terrain precision (L1TP) products registered to GCPs; Collection 2 tied to Sentinel-2 GRI-aligned ground control (USGS docs, unverified)one imagedirect'There are no restrictions on Landsat data downloaded from the USGS; it can be used or redistributed as desired' (request source statement)opentruth to make
30 m provider-registered archive (AWS copy requester-pays, Planetary Computer anonymous): known warp; low
40-year archive gives extreme temporal gaps.Known warps; GRI GCP chips or national GCPs as check points
USGS NJ orthoimagery time series (Aligning Across Large Gaps in Time)1 m USGS orthoimages of a New Jersey suburb, spring/summer/fall of 2006-201710 images 7582x5946; pairs generated on the flyorthoimages aligned by USGS; random homographies addedboth imagesdirectUSGS public domain (source imagery)openneeds an account
a recipe rebuilt from USGS EarthExplorer, which needs an account for some products
Recipe, not a packaged set: rebuild from USGS EarthExplorer (account needed for some products).Known homography/smooth warp on aligned orthos
VEDAI (Vehicle Detection in Aerial Imagery)aerial orthophotos (Utah AGRC), RGB and near-infrared channels, 12.5 cm1,210 RGB/NIR image pairs 1024x1024 (and 512x512)channels from one acquisition, aligned by construction; vehicle boxesone imagedirectresearch use (unverified)unknowntruth to make
RGB and NIR of one acquisition: known warp; research-use terms, original server not confirmed; after the CoMIR aerial set
Primary page not confirmed (original Caen server); aligned by construction only.Known warp between RGB and NIR channel gives exact multimodal truth
VPAIR (aerial visual place recognition)light-aircraft nadir images (>300 m altitude) vs rendered reference from public geodata (orthophoto + DEM), Bonn to Eifel2,706 query-reference pairs + 10,000 distractors6-DoF reference poses (GPS + attitude) and dense depth for rendered referencesone imagerequestacademic use only; commercial use prohibitednon-commercialneeds a request
the data needs a request or a data use agreement
Form needed for full data.Render reference at the query pose, refine with 10 manual points per pair
WHU Building change detection subset (Christchurch 2012/2016)aerial orthophotos of Christchurch, Apr 2012 vs 2016 (post-earthquake rebuild), RGB 0.2-0.3 mone large pair 32,507 x 15,354 px (20.5 km2); usually tiled 256x2562012 image geo-rectified to 2016 with 30 manually selected ground GCPs, 1.6 px accuracy (dataset text); building vectors for both datesboth imagesdirectnone stated on page (copyright GPCV Lab); source imagery from LINZ (CC BY 4.0 at LINZ, unverified)unknowntruth to make
one large pair already rectified on 30 manual GCPs (1.6 px): a known warp by blocks; one scene; low
Single large co-registered pair with stated residual (1.6 px); building polygons on both dates allow corner-based check points; LINZ originals at 7.5 cm are even larger.Building corners from the two vector layers as hundreds of point pairs; or known smooth warp on the full-resolution pair processed chunk by chunk
WHU-OPT-SARGaoFen-1 optical (RGB+NIR, 5 m) and GaoFen-3 SAR over Hubei, China, same scene100 image pairs 5556x3704None for registration (pre-aligned, method not stated); pixel land-use labels shared by both imagesone imagedirectNot statedunknowntruth to make
100 pre-aligned GF-1/GF-3 scenes of 5556x3704: known warp; no licence; after Major TOM
Fails truth; built for land-use classification.Apply known smooth warps to the SAR image (exact dense truth); check base alignment with 20 points on 10 scenes
WHU-SEN-CitySentinel-1 SAR to Sentinel-2 optical, 10 m, Chinese cities~18.5K pairs 256x256 (3MOS table)co-registered by construction (geocoded Sentinel products); method not describedone imagedirectnot seen (unverified)unknowntruth to make
S1/S2 city pairs on one grid: known warp; licence not seen; after Major TOM
Made for SAR-to-optical translation. status change: Round 4 count: listed in 2 of the 38 round-4 surveys and benchmark papers (Remote sensing registration review 2026; SAR-optical review 2025 (MultiResSAR)).Known warps; check points on test subset
WHU-StereoGaoFen-7 satellite stereo pairs over Chinese cities, epipolar rectifiedtrain/test epipolar pairs with disparity (count in paper, unverified)disparity maps from airborne LiDARone imagedirectnot stated (unverified)unknownwaits for a tool
epipolar stereo with LiDAR disparity: a 1D task with occlusions, outside the smooth-field contract; waits for a stereo disparity scorer
1D epipolar task.Project the LiDAR DSM through refined RPCs into both views for 2D correspondences beyond epipolar disparity
WHU-TLC (ZiYuan-3 three-line camera)ZY-3 forward/nadir/backward panchromatic images with RPCs, multi-view satellite stereotriplets of 5120x5120 crops (count unverified)DSM ground truth from LiDAR/stereo; RPCsone imagedirectunknownunknowntruth to make
ZY-3 triplets with RPC and DSM (page not confirmed): dense correspondences by projection; low
Primary page not confirmed in this sweep.DSM + RPC projection gives dense correspondences between views
xBD / xView2 (pre/post disaster)Maxar WorldView pre- and post-disaster images, 19 events, 0.3-0.8 m~11,000 pre/post pairs 1024x1024 (22,068 images)pairs geo-registered by Maxar metadata, known residual misalignment from off-nadir; building polygons on both datesboth imagesaccountCC BY-NC-SA 4.0 (xView2 terms, unverified)non-commercialneeds an account
the download needs an account
Polygons drawn per date give paired building outlines, i.e. correspondences. URL is the dataset's GitHub (homepage links from there). Data via https://xview2.org/ (login).Polygon corners of the same building in pre and post images as point pairs; check 20 pairs manually
Yuanxin Ye multimodal test pairs (HOPC, CFOG)Visible-infrared, optical-LiDAR (depth and intensity), optical-SAR, optical-map; remote sensingAbout a dozen pairs across 4 typesControl/check point files per pair (CP.txt), picked by the authorsboth imagesdirectCode: provided for research purposes only. Any commercial use is prohibited (patent). Data: not statedunknowntruth to make
about ten small multimodal pairs whose point files seed the matcher (initial correspondences), not a held-out truth: people points; research-only terms; low
Tiny, but real points on real multimodal pairs. Same files in yeyuanxin110/HOPC.Second person re-clicks the points to get observer error; add 10 more pairs per type

Non-medical · 193

DatasetProblemPairsChecked againstAccessLicenceStatusNotesTo make the truth
12-Scenes (Stanford relocalisation)iPad RGB frames (2D) against an RGB-D reconstruction (3D) of 12 rooms in 4 indoor scenes12 rooms, several sequences each, train/test splitcamera poses from global bundle adjustment of all sequences of a scene (Structure.io + iPad, BundleFusion-style)both imagesdirectCC BY-NC-SA 4.0non-commercialwaits for a tool
iPad frames against RGB-D reconstructions of rooms (camera pose): waits for a 2D/3D pose error
real; default intrinsics, no calibrationsame as 7-Scenes: second pGT from SfM, or manual 2D-3D clicks on a test subset
2D DIC Challenge 1.02D speckle images of flat surfaces, same surface deformedabout 15 samples, each a short image series (rigid sub-pixel shifts, sinusoidal fields, experimental tests)imposed displacement for synthetic samples (e.g. Samples 1 and 5 rigid 0.05 and 0.1 px steps, Sample 14 sine of rising frequency made with TexGen and FFT warping); none for experimental samplesboth imagesdirectnot stated on the iDICs pageaskin the corpuspriority 3
In the corpus, ranked apart: nine test pairs, one a speckle pattern with a documented displacement (Samples 1 to 7 shifts, 14 sine, 15 strain concentration), truth synthetic; rotations, Fazzini gradients and experiments left out; no licence stated (ask).
Standard of the DIC community; sub-pixel accuracy test, very small displacements (0.1 px amplitude on Sample 14).For the experimental samples, run 2-3 independent DIC codes and keep only points where they agree, as a weak consensus check.
2D DIC Challenge 2.0 (Star images)synthetic 2D speckle images, same pattern deformedStar image series (noiseless and noisy, several sizes)imposed vertical sinusoidal displacement whose frequency rises across the image, exact; made by B. Blaysatboth imagesdirectnot stated on the iDICs pageunknownleft outpriority 3
one speckle pattern in six noise and strain variants: a resolution test, not independent pairs
Measures spatial resolution versus noise; a ready resolution test for any 2D deformable method.
7-Scenes (Microsoft RGB-D relocalisation)Handheld Kinect RGB-D frames of 7 indoor scenes to a dense TSDF reconstruction; image-to-3D-model pose recovery7 scenes, several sequences of 500-1000 frames eachCamera-to-world pose per frame from KinectFusion tracking (method output)both imagesdirectNon-commercial use only (licence agreement on download)non-commercialwaits for a tool
Kinect RGB-D frames against a TSDF model, KinectFusion poses: waits for a 2D/3D pose error
Non-medical; poses are KinectFusion output with known drift; 2D-to-3D model registration analogueUse the refined poses published by later relocalisation work (e.g. SfM-refined 7-Scenes poses) and report error against both
A2D2 (Audi Autonomous Driving Dataset)6 cameras (2D) against 5 LiDARs (3D), German drivingabout 41,000 labelled frames plus 390,000 unlabelledcalibrated camera-LiDAR extrinsicsboth imagesdirectCC BY-ND 4.0 (per memory; site did not answer)askwaits for a tool
camera images against LiDAR point clouds (calibrated extrinsics): waits for a 2D/3D pose error
real; a2d2.audi did not respond to a status check (unverified availability)
Aachen Day-Night (v1.0 and v1.1)phone photos by day and night (2D) against an SfM map (3D) of Aachen old town; long-term localisationv1.1: 6,697 reference images, 1,015 queries (824 day, 191 night)query poses from SfM with manually added 2D-3D matches for night images (hidden on the evaluation server)both imageshiddenterms in each dataset readme (not read; unverified)unknownleft out
its images or its truth are not released
real; test poses withheld behind the server; day-night changes make it a strong cross-condition test status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Image matching survey 2025).Hand points on day-night pairs drawn from the public reference images, or reprojected SfM tracks as pseudo truth.
AIHub riparian change detection, Han and Nakdong rivers (Korea, datasets 71643 and 71644)bitemporal aerial or satellite images of river zones, same areaunverifiedchange labels; images co-registeredaccountAIHub terms (eligibility for non-Korean users unverified)askneeds an account
the download needs an account
change detection, so pairs are aligned by constructionapply known warps for exact truth
Ailsa Craig micro-granite triaxial 4D uCT (Cartwright-Taylor et al., BGS)micro-granite (fresh and heat-treated) loaded to failure at 15 MPa confinement, synchrotron uCT at each stress step (SOLEIL PSICHE)two time series (counts unverified)noneone imagedirect"Licence: Not set" (data.gov.uk)asktruth to make
3D synchrotron micro-CT stress steps of two granite samples (via BGS NGDC), no truth; points by hand on the near-rigid early steps would make it; reuse needs BGS permission
Access route through BGS/NERC not checked.Early stress steps near-rigid; crack network as a discontinuity; 20 manual points
Aluminium pin corrosion 4D synchrotron CT (Bellos et al., Diamond I13-2)salt-water droplet on an aluminium pin, synchrotron CT time series, corrosion21 undersampled tomograms (91 projections each)partial manual 4D segmentation on a few slices (zenodo 5564782); series already alignedone imagedirectCreative Commons Attribution 4.0 Internationalopentruth to make
21 reconstructed 3D tomograms in one 193 GB 7z archive, already aligned, only partial slice labels; zero motion on the pin or a known warp would make it
Mostly a zero-motion test with appearance change and heavy undersampling noise.Zero-displacement truth on the pin; add known smooth warps
AM Bench 2022 IN718: X-ray CT + serial-section EBSD (NIST)IN718 PBF-LB leg, X-ray CT vs serial sectioning EBSD, plus build file and thermography, same specimen, 3D multimodalone region of interest, several modalitiesspatially registered by NIST (method not detailed on the page)directhttps://www.nist.gov/open/licenseopentruth to make
3D X-ray CT and serial-section EBSD, optical and BSE volumes of one region (DREAM3D), registered by NIST into build coordinates; pores matched by hand would make it
Rare 3D multimodal materials set; registration by the makers.Pores visible in both CT and EBSD as fiducials: match 20 pores by hand
Archimedes Palimpsest digital release10th-century palimpsest folios, multispectral bands, pseudocolour, X-ray fluorescence maps and Heiberg 1906 photos, same leafevery folio: many bands; XRF or Heiberg images for some folios; 2016 supplement 22 folio sides x 16 bandsper-folio registered image sets made by the project ('registered Heiberg images and registered XRF images'); transcriptions spatially mapped to all registered imagesboth imagesdirectCreative Commons Attribution 3.0 Unportedopentruth to make
XRF, Heiberg-photo and multispectral images of each folio come registered by the project with no stated accuracy (CC BY 3.0, per-folio download): two people marking about 20 points on ten XRF-optical folios would make a multimodal truth
Registration by the makers, accuracy not stated. Bands share a camera; XRF-to-optical and Heiberg-photo-to-optical are the real multimodal tasks.Use mapped transcription word boxes as points; or click 20 points on 10 XRF vs optical pairs
Arctic sea ice motion: RADARSAT-2 ScanSAR with GNSS buoys (SIDEx)SAR images of drifting sea ice a few days apart, same area1 raw SAR image pair (PNG) plus flows of 48 modelsGNSS buoy positions projected into pixel coordinatesboth imagesdirect: Zenodo, no accountCC BY 4.0 (Zenodo)openleft out
The Zenodo record (CC BY 4.0, 13.8 GB) holds one raw SAR pair as PNG (2021-03-22 and 2021-03-25), the flows of 48 models (13.8 GB zip) and buoy positions in pixel coordinates for 54 scene dates (about 32 soft buoys plus GPS buoys a scene). Only the one pair has images; the other RADARSAT-2 scenes are not deposited, so one case with some 30 points.
physical markers on the ice; only one raw pair is depositedadd more SIDEx RADARSAT-2 scenes with the same buoy tracks
Argoverse 2 (sensor dataset)7 ring cameras and 2 stereo cameras (2D) against two 32-beam LiDARs and HD maps (3D), 6 US cities1,000 sensor scenes; 20,000 unlabelled LiDAR sequencesfactory camera-LiDAR calibration and city-frame poses; I2P truth = calibration plus synthetic perturbationboth imagesdirectCC BY-NC-SA 4.0 (data); code MIT (unverified)non-commercialwaits for a tool
camera images against LiDAR sweeps and HD maps: waits for a 2D/3D pose error
realHand-check non-box movers on a sample of sweeps.
astroalign simulated star-field benchmarksimulated star fields, same field with rotation, scale and shift, different PSF and noise640 pairs generated in the paper (not stored); any number can be generatedknown similarity transform by construction (Beroiz, Cabral, Sanchez 2020)both imagesdirectMIT (code)openleft out
no images released: the paper's 640 simulated star-field pairs were not stored (only timing CSVs)
Synthetic and linear only; only timing CSVs are in the repo. Fails 'real images'.Simulate with known non-linear distortion and PSF change; better, use real frames with Gaia control points (see ZTF/HST entries)
AutoFlowsynthetic layered 2D shapes with learned rendering settings40,000 training pairs (static set)exact dense flow by construction (Sun et al., Google)both imagesdirect: storage.googleapis.com/autoflow, no account; zip parts readable by rangesCC BY 4.0 (project page)openleft out
Google's static AutoFlow set (four zip parts on an open Google Cloud bucket, about 96 GB, 'License: CC-BY' on the project page) is a pretraining set of 40,000 rendered pairs of abstract layered shapes with exact flow by construction. No image in it is a real scene, and its layers move over each other with occlusions that a one-to-one registration field cannot represent; the corpus's own known warps on real images already give exact synthetic truths.
Training set only, abstract shapes.
BAM AM gyroid lattices: XCT vs CAD (Chand et al. 2025)AlSi10Mg gyroid lattices, XCT (5 um) vs voxelised CAD, nominal-actual 3D7 samplesmanual rigid alignment to CAD on the base plate (VGStudio); no deformation truth; Dice after binarisingone imagerequestno data licence stated ('available upon request')askneeds a request
the data needs a request or a data use agreement
Small; truth is overlap only.Simulate shrinkage on CAD with a known field; add fiducial spheres to the build plate in new builds
BAMBI matched RGB-thermal UAV wildlife frames (Zenodo 19034999)nadir drone RGB vs thermal frames of forest wildlife5,305 matched pairs13,057 co-registered bounding boxes (authors' alignment)both imagesdirectCC BY 4.0opentruth to make
RGB and thermal frames aligned by the authors (CC BY 4.0, 18.7 GB): a known warp on the matched pairs, or points by two people to check the alignment, would make the truth
alignment made by the authorstwo people mark 10 points per pair on 20 pairs to check the given alignment
Battery Imaging Library: operando micro-CT (Imperial College, Diamond I12)commercial Li-ion, LiFeS2 and alkaline Zn/MnO2 cells during cycling, synchrotron micro-CT, same cell over timemany operando datasets (e.g. Datasets 3, 7, 15, 37, 52 on Zenodo)noneone imagedirectZenodo records: cc-by-4.0; library text: academic and educational purposes, commercial use to be discussed with the authorsaskwaits for a tool
the operando micro-CT records hold raw projections only, no reconstructed volumes, and no truth: waits for a tomographic reconstruction step
Raw data, reconstruct yourself. Electrode swelling and gas give real non-rigid motion.Casing and tabs rigid: rigid truth; DVC inside checked by hand on 20 particle centroids
Bentheim sandstone micro-CT at 1, 20 and 30 MPa (Lebedev)dry Bentheim sandstone in a pressure cell, micro-CT at three confining pressures and two voxel sizes, same sample3 pressure states x 2 resolutions (1.59 and 3.35 um)none; elastic deformation is tiny so pairs are near-rigidone imagedirectCreative Commons Attribution 4.0 Internationalopentruth to make
3D micro-CT of one sandstone at three pressures and two voxel sizes, near-rigid, no truth; grain contacts by hand would make it
Good repeat-scan and cross-resolution test.Near-zero-displacement truth between pressure steps (rigid only) and known scale between resolutions; check 20 grain contacts by hand
Benthic long-term visual localisation dataset (AUV, up to 6 years)near-nadir AUV stereo imagery of the same seafloor sites revisited over years5 reference sitessub-decimetre registered camera poses and footprint-based ground truthboth imagesdirectunverifiedunknowntruth to make
Truth is registered camera poses and overlap links, not pixel correspondences (stereo depth plus poses would give them); the paper's page names no download location or data licence
Seafloor change between visits.
BioID Face Database (with FGNet 20-point markup)2D grey-level frontal face photographs, 23 people, between images1,521 imagesmanual eye positions; FGNet markup scheme adds 20 facial landmarks per imageboth imagesdirect'Free to use for face detection research purposes ... without further permission'askneeds a request
Direct download, but use is granted for face-detection research only: registration needs BioID's permission; 1,521 photos of 23 people with hand-placed FGNet points (privacy review advised)
Use is limited to face-detection research by the wording; registration benchmark use needs a check. People's faces: privacy review advised.
BOP-Classic (LM, LM-O, YCB-V, HB, IC-BIN, TUD-L, TYO-L)RGB-D photos (2D) of household and toy objects against textured 3D models (3D); 6-DoF pose7 datasets, 2-33 objects each, thousands of test images6D poses from marker boards, manual alignment or ICP refinement, depending on datasetboth imagesdirectLM CC BY 4.0; LM-O, TUD-L CC BY-SA 4.0; YCB-V MIT; HB CC0; TYO-L CC BY-NC 4.0; IC-BIN not statedopenwaits for a tool
RGB-D photos against textured object models (6D pose): waits for a 2D/3D pose error
real; HB test poses hidden; YCB-V has known erroneous GT on some frames (BOP subset filters them)
C3 (floor plan to photo correspondences, C3Po)ground-level photos and building floor plans of the same scenes90K plan-photo pairs, 597 scenes, 85K camera poses153M pixel correspondences from SfM reconstructions manually registered to plansboth imagesdirectunverifiedunknownleft out
A floor plan and a perspective photo are related by a many-to-one projection (walls collapse onto lines), not a field a registration can return; licence unverified
Cross-modal 2D-2D.
Cai PIV dataset (Problem Class 1 and 2)synthetic 2D particle image pairs (fluid flow)over 10,000 pairs at 256x256 (uniform, back-step, cylinder, DNS turbulence, JHTDB, sea surface flow)dense flow field from CFD and turbulence databases used to render the particles; syntheticboth imagesaccount: both Google Drive folders of the README ask to sign in (checked 2026-10-26), Baidu Pan needs an accountprovided for research purposes only, all rights reserved, commercial use requires consentaskneeds an accountpriority 3
The data sit in two Google Drive folders that now ask to sign in, and on Baidu Pan, which needs an account; no public mirror found. Research use only (ask).
Standard learned-PIV training and test set; images are sparse particles, unlike speckle or anatomy. FlowData_test.list holds 449 pairs from 16 flow sources (backstep at 4 Reynolds numbers, cylinder at 5, DNS turbulence, JHTDB channel and channel_hd, isotropic, MHD, SQG, uniform): one pair a source would give about 15 independent cases.
Callovo-Oxfordian claystone x-ray and neutron time series (Stavropoulou 2020)simultaneous x-ray and neutron CT of claystone during water uptake (swelling, cracking)time series in both modalities, neutron already registered to x-raynone (author registration and DVC only)one imagedirectcc-by-4.0opentruth to make
3D x-ray and neutron CT time series of swelling claystone, neutron registered to x-ray by the authors, no truth; a known transform or crack landmarks would make it
Fails truth. Real multimodal 3D pair plus deformation over time.Unregister by a known transform to test x-ray/neutron registration; sparse crack-tip landmarks for the swelling series.
Cambridge LandmarksOutdoor phone video frames of Cambridge landmarks to an SfM 3D reconstruction; image-to-model pose recovery6 scenes, thousands of frames (King's College record linked)SfM (VisualSFM) camera poses, method outputboth imagesdirectCC BY-NC-SA 2.0 UKnon-commercialwaits for a tool
phone video frames against an SfM model: waits for a 2D/3D pose error
Non-medical; each scene is a separate repository recordRe-run a modern SfM (COLMAP) on all frames and keep frames where both pose sets agree within a few cm
Carbonate dolomite-calcite dissolution, 40 segmented micro-CT volumes (NERC/BGS)dolomite-calcite carbonate under CO2-acidified brine at reservoir conditions, micro-CT time series, same sample40 volumes, 33 min apart, 5.2 um voxelsnone; 3-phase segmentations (pore, dolomite, calcite) per time step by the makersdirectunverifiedunknowntruth to make
40 segmented 3D micro-CT volumes (labels only, no grey levels) of one carbonate during dissolution, OGL via BGS NGDC, no truth; unreacted grains would make it
Segmented images only (8-bit raw), not grey levels; access terms unverified.Unreacted grains give label-overlap truth; slow-reacting dolomite gives a near-zero-motion check
CCPi synthetic magma DVC datasetsynchrotron CT (Diamond I12) of synthetic magma, before and after compression1 pair (dataset_0 no load, dataset_1 compressed)noneone imagedirectcc-by-4.0opentruth to make
one pair of 3D synchrotron CT volumes (1520x1257x1260, before and after compression), no truth; a known warp or matched bubbles would make it
Fails truth. Large single pair, NumPy and MetaImage headers.Apply a known smooth warp to dataset_0 for exact truth; for the real pair, bubbles matched by a program and checked by a person.
Challenging Data for Stereo and Optical Flow (HCI 2012)outdoor car-mounted stereo sequences, frame to frame and left to right11 sequencesnone (unverified)directnone on recordunknowntruth to make
Real car-mounted stereo sequences with no truth and no licence: points by two people on ten frame pairs would make one
Heidelberg HCI set (Meister, Jähne, Kondermann 2012); HD1K from the same group is already listedknown warps only; sparse manual point tracks by two people on a few frames
Closer to Van Eyck (Ghent Altarpiece and VERONA images)Ghent Altarpiece and other Van Eyck paintings, visible, infrared, IRR, X-ray and macro, before, during and after treatmentevery panel in several modalities and treatment stages (about 100 gigapixels)stitched and multimodally registered by KIK-IRPA with UGent/VUB/imec for the viewer; method and accuracy not published on the siterequestunverified; site has a 'Using images from this site' page; KIK-IRPA says VERONA images are 'available online for free for research purposes'askneeds a request
the data needs a request or a data use agreement
Viewer-based; bulk full-resolution download unverified. Treatment stages give same-modality temporal pairs too.Two people click 20 craquelure points on 10 VIS-IRR and VIS-XR pairs; spot-check the site's own registration
CO2XCT (supercritical CO2-brine coreflood micro-CT)micro-CT of the same rock core at successive flooding stages under pressureseveral stages, 1,700 slices each at 25 umnoneone imagedirectunverifiedunknowntruth to make
3D micro-CT of one core at successive flooding stages (raw and segmented, direct download), no truth; grain centroids matched and checked would make it
Sci Data 2026. Core held in a fixed holder, so motion is small.grain centroids matched by a program and checked by a person
Concrete in situ x-ray compression (Stamati 2020)x-ray CT of concrete under uniaxial and triaxial compression (5, 10, 15 MPa)4 load series plus a double scan of the reference (doubleScanUniaxial)zero motion between the double scans; otherwise only author DVCone imagedirectcc-by-4.0opentruth to make
3D x-ray CT load series of concrete, only a zero-motion double scan and the authors' DVC; tracking the aggregates labelled in the reference scan would make it
Labelled aggregates provided for the reference scan.Aggregate centroids tracked by a program and checked by a person; double scan as zero-motion truth.
Content-aware unsupervised deep homography dataset (CA-UDHN, Zhang 2020)real small-baseline video frame pairs: regular, low texture, low light, small and large foreground80k pairs; 4.2k test pairs6-8 hand-marked matching points per test pair (authors, UESTC/Megvii); v2 coordinates refined with feature pre-matchingboth imagesdirectMIT (repository licence); no separate data licenceunknowntruth to make
6-8 hand points a test pair, placed for one homography, so an affine alignment absorbs most of the 4-8 px no-warp error: ten or more points over all depths on 20 pairs with parallax or moving foreground would make it; research use only
Real scenes break a single homography (parallax, moving objects); Google Drive/Baidu download.Two annotators, 20 points per pair on a subset, to measure annotation noise.
CryoBench (synthetic cryo-EM heterogeneity benchmark)simulated cryo-EM particle images (2D projections) of known 3D density maps; pose and conformation recovery5 datasets (IgG-1D, IgG-RL, Ribosembly, Tomotwin-100, Spike-MD), 100,000-335,000 particles eachexact poses, CTFs and conformation labels by construction (simulation from atomic models)both imagesdirectCC BY 4.0openwaits for a tool
simulated 2D particle projections against 3D maps with exact poses: waits for a 2D/3D pose error
synthetic; projection-to-volume registration is the core task (pose estimation)
DCCRP (dual-camera photogrammetry of slope markers with RTK)dual 4K time-lapse cameras viewing markers on a natural slope, day and night, controlled displacements14 markers over 48 h; 13 markers over 5 displacement stagesRTK GNSS 3D coordinates of markers (5 repeats per stage)both imagesdirectCC BY 4.0openleft out
Only the survey markers move, by hand, on a still slope: no deformation between the images to score
Markers already in the images; image-based displacement vs survey truth.
DIC_noise_bias (Baldi)2D camera images of a speckled target translated by a piezo stage under a telecentric lensseveral rigid translation series at noise levels from good to very weak lightknown sub-pixel translation from the PZT x-y stage; real imagesboth imagesdirect: Mendeley Data public files (589 files), no accountCC BY 4.0 (Mendeley Data)openleft out
12 series of 46 images (three light levels by four averaging levels, raw .bin, 1.06 GB) of one speckled target moved by a piezo x-y stage under a telecentric lens: the motion is a known rigid translation, which the corpus's rigid alignment removes, and one target cannot give independent test pairs. A known warp over two real shots would make a 2D noise test, ranked apart, if one is wanted; CC BY 4.0 on Mendeley Data.
Rigid only; good real-noise sub-pixel bias test.
Digital Rocks DRP-221/264/319: solute transport micro-CT (Ghent UGCT)sintered glass, Bentheimer sandstone and Savonnieres limestone, tracer injection, fast lab micro-CT, same samplea scan every 12-15 s plus a slow high-quality pre-scan per experimentreconstructed sets cropped and registered by the makers (Van Offenwert et al. 2019); rock staticdirectODC-BY 1.0 (raw project 319)opentruth to make
3D micro-CT tracer series of static rock (fast scans and a slow pre-scan), registered by the makers, no truth; zero motion or a known warp would make it; public archive zips
Moved to the Digital Porous Media Portal.Zero-motion truth; fast vs slow scan pairs test registration under noise and quality mismatch
Digital Rocks DRP-378: 4D oil and water flood in Luxembourg sandstonetwo twin Luxembourg sandstone samples (water-wet and mixed-wet), lab micro-CT (TESCAN DynaTOM) during oil and water floodstime-resolved series per sample (counts unverified)none; rock frame is staticone imagedirectODC-BY 1.0opentruth to make
3D micro-CT flood series of two sandstone samples with a static rock frame, no truth; zero motion on the frame would make it; public 88 GB archive
Status change: Digital Rocks Portal is now the Digital Porous Media Portal (DOIs redirect there); download terms there unverified.Zero-displacement truth on the rock frame with fluid-driven appearance change
DIR300real phone photos of warped documents vs their flat images300 pairsflat image only (Feng et al., USTC)one imagedirectnone stated (repository licence NOASSERTION)unknowntruth to make
300 phone photos of warped documents against their flat images, no correspondences, licence NOASSERTION: points by two people (or DocAligner flow checked by them) would make it
Same metric weakness as DocUNet; Google Drive download.Dense flow by DocAligner, checked with 20 hand points on 30 pairs.
Discontinuity DIC Challenge (iDICs)2D DIC images with cracks and slip (announced)not releasedhiddennot statedunknownleft out
its images or its truth are not released
Announced on idics.org as 'more details to come'. Discontinuous fields are a gap in registration benchmarks.Synthetic images with an imposed crack opening (exact), as the DIC community plans.
Doc3Dsynthetic renders of real captured paper shapes with document textures100k images, each with its flat sourceexact backward map, UV and 3D coordinates by rendering (Das et al., Stony Brook)both imagesaccountMIT (repository and Hugging Face card); textures include Yes! Magazine 'under Creative Commons Licence'openneeds an account
the download needs an account
Hugging Face gated (login); meshes by email; old route needed a Google form.
DocAligner real-virtual pairs (+ DocAlign12K)real photos of printed documents paired with the clean digital page~2.5k flow-annotated real pairs, 800 illumination pairs from WarpDoc, DocAlign12K synthetic (12k)dense flow predicted by DocAligner and used as labels (Zhang et al., SCUT); synthetic flow exactboth imagesdirectnone stated (no repository licence)unknowntruth to make
The real pairs' flow is DocAligner's own output (people checking it with points would make it semi-automatic); DocAlign12K is synthetic; no repository licence; its 800 illumination pairs come from WarpDoc
Real-pair truth is model output, not independent.Two people place 20 points on 30 pairs to measure flow error.
DocUNet benchmarkreal photos of bent and folded paper documents vs their flatbed scans130 photosflat scan per document only; no pixel correspondence (Ma et al., CVPR 2018)one imagedirectnone stated on the pageunknowntruth to make
130 photos of 65 folded documents against flatbed scans, no correspondences (its metric runs SIFT Flow): points by two people would make the truth; no licence stated
Fails strict criterion: metric relies on SIFT Flow. 64_1/64_2 rotated 180 degrees (DocGeoNet note).Print documents with a dot grid or AprilTags before warping; or DocAligner flow checked with hand points.
DSEC optical flowreal driving with stereo event cameras + RGB, day and nightflow truth on training sequences; test flow hidden (benchmark)flow from lidar depth + ego-motion (Gehrig et al., UZH)both imagesdirectCreative Commons Attribution-ShareAlike 4.0 International public license (CC BY-SA 4.0) ... Using the material for commercial purposes is allowed.openleft out
Flow truth is given only on the left event camera's grid, which records events, not images; the RGB frames are not on that grid; test flow hidden
Mostly event data; moving objects less accurate.Hand-check flow on moving objects in a sample of frames.
DTU MVS 2014 (robot image data)images from 49 or 64 robot-arm positions (2D) and structured-light reference scans (3D) of tabletop objects124 scenes (80 used in evaluation), 7 lighting conditionscamera poses from a calibrated 6-axis industrial robot arm (MATLAB calibration toolbox); structured-light scans as reference geometryboth imagesdirectdescribed as freely available; no named licence foundunknownwaits for a tool
robot-arm photos against structured-light scans, poses from the calibrated arm: waits for a 2D/3D pose error
real; pose truth comes from a mechanical rig, independent of image content, which is the strongest kind
Dual-camera close-range photogrammetry of slope markers with RTK (DCCRP)4K time-lapse image sequences of a natural slope with markers, two cameraslong sequencesRTK GNSS 3D coordinates of the markers, synchronisedboth imagesdirectCC BY 4.0openleft out
same Zenodo record (21441791) as the 'DCCRP (dual-camera photogrammetry of slope markers with RTK)' row
physical markers with survey-grade positions
Dubrovnik6K and Rome16K (2D-3D localisation against SfM models)Internet photos (2D) against large SfM point clouds (3D) of Dubrovnik and RomeDubrovnik6K: about 6,000 images with 800 queries; Rome16K: about 16,000 images with 1,000 queriesquery poses from the full SfM reconstruction before the queries were removed (pseudo ground truth)both imagesdirectnot stated (unverified)unknownwaits for a tool
Internet photos against SfM point clouds, query poses from the full reconstruction: waits for a 2D/3D pose error
real; classic; truth is SfM
DVC Challenge 1.0 syntactic foam scans (in DVC Challenge 2.0 repository)micro-CT of a syntactic foam cylinder, same specimen, 6 labs with different CT systems7 stacks of 3-7 volumes (reference, 2 repeats, 2 axial 1 mm moves, 2 radial 1 mm moves)rigid translation from a precision stage, zero motion for repeat scans; real scansboth imagesdirect: NIST Public Data Repository, one zip per stack (DVC1_S1 to S6, about 19.3 GB), no account (file server unreachable at the time of the check)NIST open licence (record's license field https://www.nist.gov/open/license); license.txt of the bundle not readopenleft out
Seven XCT stacks of a syntactic foam (glass microballoons in elastomer) from seven systems in six labs, 3 to 7 volumes each (350 to 996 voxels a side, 7 to 18 um): repeat scans and 1 mm axial or radial stage moves (preprint Table 2). The motion is rigid, which the corpus's own rigid alignment removes, and the specimens are one material; a known warp between two repeat scans would make a 3D noise test ranked apart if wanted. The NIST record states the NIST open licence; the per-dataset license.txt could not be read (the NIST file server timed out on 2026-10-13).
Rigid only. Measures CT distortion and noise floor, not deformable accuracy. Sizes 350-996 voxels a side.Apply a known smooth warp to one repeat scan and pair it with the other repeat: real noise, exact non-rigid truth.
DVC Challenge 2.0 (NIST, iDICs)Digital volume correlation image pairs and series: lab and synchrotron micro-CT, neutron CT, confocal; foams, polymers, concrete, sand, soil, sandstone, beads, synthetic; same object before and after load37 datasets from 19 contributors (pairs and loading series)Synthetic sets made with a released generator (known displacement, unverified per set); DVC Round Robin 1.0 repeat scans (Croom 2021); experimental sets without truthdirectNIST open licence (public domain style)openleft out
the whole NIST repository, already sorted in its five 'DVC Challenge' subset rows (1.0 foam, synthetic beads, fibre and wood, experimental microscopy, experimental XCT and neutron)
Non-medical but exactly the dense-displacement problem; README and data_summary list per-set truth (not read: server timed out).Apply known smooth or FE-derived displacement to the unloaded scans chunk by chunk; repeat scans give zero-motion truth; rigid translations give exact truth
DVC Challenge 2.0 experimental microscopy and flow setsmultiphoton and confocal stacks of hydrogels and collagen gels under tension, indentation and inclusion contraction; scanning PIV volumes of swimming brine shrimp and jellyfish; SHG of human lamina cribrosa under pressureFranck 28 volumes, Yang hydrogel 2, Notbohm S1 5, Fu shrimp 26+26, jellyfish 2, Nguyen lamina 5none beyond noise floor (embedded beads give trackable points)one imagedirectNIST open licence (record points to https://www.nist.gov/open/license; per-dataset license.txt not read)opentruth to make
3D multiphoton, confocal and scanning-PIV stacks under load, no truth beyond the noise floor; embedded beads tracked and checked would make it
Anisotropic voxels (z 2-5x coarser). Lamina cribrosa is human eye tissue, so not strictly non-medical.Bead centroids detected in both volumes and matched by tracking (as in TFM work) give dense sparse-point truth; verify a sample by eye.
DVC Challenge 2.0 experimental XCT and neutron setslab and synchrotron micro-CT and neutron CT of foams, 3D-printed polymer lattices, sand, soil, sandstone, concrete under load; same specimen over load steps16 series, 3-16 volumes each (Landauer and Sriram foams, Kafka and Shaikeea polymers, SPAM granular x-ray and neutron)none beyond noise floor and rigid-body steps; DVC fields by contributors are not truthone imagedirectNIST open licence (record points to https://www.nist.gov/open/license; per-dataset license.txt not read)opentruth to make
3D lab, synchrotron and neutron CT load series (volumes up to 1470x1379x2000), no truth; tracked lattice nodes, grains or pores checked by a person would make it
Real large-deformation 3D series up to 1470x1379x2000 voxels: good stress test for out-of-core registration.Lattice nodes, sand grain centres or foam pores tracked by a program and checked by a person give sparse 3D landmarks on both volumes.
DVC Challenge 2.0 synthetic bead volumes (Patel, Yang)synthetic fluorescent-bead volumes mimicking confocal images, same volume deformedPatel: 2 volumes (sinusoidal field); Yang S1-S3: 12/7/7 volumes (translation, stretch, rotation)imposed analytic displacement, exact by construction; syntheticboth imagesdirect: NIST Public Data Repository, synthetic_data/Patel_Beads.zip and Yang_Beads_S1-S3.zip (about 2.5 GB), no account (file server unreachable at the time of the check)NIST open licence (record); license.txt of the bundle not readopenleft out
Synthetic confocal-like bead volumes (512 x 512 x 192, 8 bit) made by Poisson-disc sampling and a PSF: Patel, 2 volumes under a sinusoidal displacement; Yang S1 to S3, 12, 7 and 7 volumes under translations, uniaxial stretches (1.0 to 1.3) and rotations (0 to 25 degrees) of one reference (preprint Table 9). Translations, rotations and stretches are linear and leave nothing after alignment, one sinusoidal pair remains, and the generator is public (GitHub), so the corpus would do better drawing its own warp on any volume.
512x512x192 voxels. Whether the field is shipped as a file or only described is unverified.
DVC Challenge 2.0 synthetic fibre network and wood models (Notbohm S2, Chen)synthetic collagen fibre network (confocal-like) under shear; numerical wood microstructures (balsa, birch, spruce)Notbohm S2: 3 volumes (5% and 30% global shear); Chen: 1 volume each, no deformationimposed global shear for Notbohm S2; none for Chen wood (single volume)both imagesdirect: NIST Public Data Repository, synthetic_data/Notbohm_Fibrous_S2.zip and Chen_*.zip (about 0.21 GB), no account (file server unreachable at the time of the check)NIST open licence (record); license.txt of the bundle not readopenleft out
Notbohm Fibrous S2 is one simulated collagen network imaged synthetically (383 x 383 x 133) in 3 states, undeformed and under 5 % and 30 % global shear, with finite-element node displacements as truth; the Chen wood models (balsa, birch, spruce, 512 cubed) are single numerical volumes with no deformation (preprint Tables 5 and 12). One network and four unpaired volumes, with the network and image generators public on GitHub: nothing a known warp drawn by the corpus would not give.
Anisotropic, sparse texture: hard for intensity methods. Wood sets need a warp applied to be usable.Apply a known smooth or orthotropic warp to the Chen wood volumes; exact truth on a fibre texture.
DVC model material: polypropylene with copper particles (Wantz)lab CT of a polypropylene specimen with embedded copper particles, moved between scansa translation series and a rotation series (number of steps unverified)known rigid translation and rotation from the setup (amounts unverified); real scansboth imagesdirect: Zenodo record 15614922, two zips of TIFF slices (readable by ranges)CC BY 4.0 (Zenodo)openleft out
Zenodo (CC BY 4.0, 35.7 GB of TIFF slices, a one-line description): one polypropylene specimen with copper particles, a reference volume and one displaced volume (Translation.zip, folder Dep_02_vx, presumably 2 voxels) plus a reference and four rotated volumes (Rotation.zip, rot_240p to rot_1920p, amounts not stated). Five rigid moves of one object leave nothing deformable to score, and each volume is about 1265 slices of 6 MB.
Built as a DVC benchmark; description on Zenodo is one line, details unverified.Copper particle centroids give sparse landmarks to check the commanded motion.
Dynamic Replicasynthetic stereo videos of moving humans and animals in indoor scenes524 videos, 145,200 stereo framesexact depth, optical flow, long-range pixel trajectories and cameras by rendering (Meta AI, Oxford)both imagesdirectDynamic Replica License Agreement: use only for 'a research, educational, or artistic purpose that is non-commercial or not-for-profit'non-commercialwaits for a tool
synthetic stereo video with exact flow and long-range tracks: waits for point tracking
Click-through licence on the project page.
egenioussBench (smartphone localisation against city mesh and LoD2)smartphone photos registered to an airborne city mesh and CityGML LoD2 model2,709 images; 412 validation with poses, 42 test with hidden posescentimetre-accurate poses from PPK and control pointsboth imagesdirectunverifiedunknownwaits for a tool
smartphone photos against a city mesh and LoD2 model: waits for a 2D/3D pose error
Image-to-3D model registration, 2026.
EgoPoints (egocentric point tracking benchmark)egocentric video with points leaving and re-entering view4.7K annotated tracks; 11K semi-real sequenceshuman-annotated tracks; automatic truth for semi-real sequencesboth imagesdirectunverifiedunknownwaits for a tool
egocentric video with annotated point tracks: waits for point tracking
arXiv Dec 2024, published 2025.
EMPIAR-10028 (Plasmodium 80S ribosome)cryo-EM particle images (2D projections) of a ribosome against its reconstructed map EMD-2660 (3D); pose estimation105,247 particles from 1,081 micrographsparticle poses from the published RELION refinement (a reconstruction, not independent)both imagesdirectEMPIAR data released under CC0 (EMPIAR policy; unverified for this entry)openwaits for a tool
2D particle projections against a 3D map: waits for a 2D/3D pose error; the poses come from the published refinement
real; truth is consensus refinement, so it rewards agreeing with RELIONsimulate particles from the deposited map with known poses and the measured CTFs (CryoBench-style) for exact truth
EMPIAR-10164 (HIV-1 Gag VLP cryo-ET tilt series)cryo-ET tilt series (2D projections) of immature HIV-1 virus-like particles against subtomogram averages EMD-3782/EMD-4015 (3D)43 tilt seriestilt alignment from gold fiducial beads; particle poses from published subtomogram averaging (reconstruction)both imagesdirectEMPIAR data under CC0 (EMPIAR policy; unverified for this entry)openwaits for a tool
2D cryo-ET tilt series against 3D averages: waits for a 2D/3D pose error; gold beads check the tilt alignment
real; fiducial beads give an independent check of tilt-series (projection-to-volume) alignmentscore markerless tilt alignment by reprojection error of the gold beads
EPFL CVLab deformable surface sequencesvideo of deforming paper and cloth vs a template image and meshseveral sequences (count not shown); Vicon cardboard/cloth capture without images3D-to-2D matches per frame to the template mesh; Varol 2012 sequences with ground truth (EPFL)both imagesdirectThis data is free of use for research purposes.non-commercialtruth to make
Deforming paper and cloth against a template whose matches are feature matches, research use only: printed markers or points by people would make a truth
Most matches are feature matches, not independent truth.Print markers on the surface, or render the Vicon meshes with texture for exact synthetic pairs.
EPFL lidar dense-flow pairs (Fountain, Herzjesu)real photos of static building facades, small and large baselineFountain and Herzjesu, small- and large-baseline pairs (a few pairs)dense optical flow and visibility from lidar scans (Strecha et al., EPFL)both imagesdirectEPFL data page: 'should only be used for research and evaluation purposes'; commercial use needs written approvalnon-commercialleft outpriority 3
two facades, four pairs: too few independent scenes
Real images with lidar dense truth; tiny and rigid.
ETH UAV dataset of nine wheat fields (raw, processed, metadata)wheat fields, close-range UAV RGB, same field up to twice a week over two seasonsnine fields, many dates (counts unverified)photogrammetric registration with fiducial ground control points; camera poses and GCP metadata (ICRA 2024 crop alignment paper)directunverified (page returned an error)unknowntruth to make
Same fields twice a week with fiducial ground control points: the GCP targets clicked by people on both photos would give control points; the record's licence could not be read
Link between this record and the ICRA paper is unverified. Close to remote sensing, which others cover.Visible GCP targets give control points; hold out GCPs to measure error
ETH3D (multi-view and two-view)indoor/outdoor scenes: high-res DSLR multi-view, low-res many-view and two-view stereo13 high-res training scenes (14-76 images each) + two-view set; test truth hiddenlaser scans registered to images, rendered depth (Schoeps et al., ETH Zurich)both imagesdirectlicensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International Licensenon-commercialin the corpuspriority 3
In the corpus, ranked apart: one pair of DSLR photographs from each of the 11 independent high-res training scenes (relief_2 shares relief's hall), 10 test and 2 development, points carried by the laser-scan depth and the provider's poses (automatic); CC BY-NC-SA 4.0. The two-view rig frames are not used.
Laser-scan truth on both images; good for wide-baseline dense correspondence. status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Image matching survey 2025).
ETH3D multi-view stereo benchmarkDSLR photos and rig video frames (2D) registered to terrestrial laser scans (3D) of indoor and outdoor scenes13 train / 12 test DSLR scenes; 5/5 rig videos; 27/20 two-view framescamera poses refined by photometric alignment of images to the laser-scan geometry (Schoeps et al. 2017); laser scan is the referenceboth imagesdirectCC BY-NC-SA (third-party listing says 3.0; eth3d.net terms not read, unverified)non-commercialsame as
Same images as the row 'ETH3D (multi-view and two-view)', under which they are built as photograph pairs (ETH3D); an image-to-laser-scan task would still wait for a 2D/3D pose error.
real; one of the few image-to-laser-scan sets with sub-pixel calibration; test geometry hiddenalready strong; for a registration task, perturb the published poses and score recovery against them
Extended CMU Seasonscar camera images across seasons (2D) against an SfM map (3D) of Pittsburgh60,937 reference, 56,613 query images; half the query poses public, half hiddenposes from SfM with manual verification across conditions (Sattler et al. 2018)both imageshiddenterms in dataset readme (unverified)unknownleft out
its images or its truth are not released
real; half the poses public, half hidden for the server
FAU historical prints registration set16th-century prints, four exemplars of each motif, between copies of the same print39 pairs (13 motifs x 3 copies to a reference)10 manual control point pairs per pair (Sindel et al. 2022)both imageshiddennot statedunknownleft out
its images or its truth are not released
Small, release not stated; wear and print differences make it a between-copies problem.Add 20 points per pair from a second annotator
FAU multi-modal panel painting dataset (CraquelureNet line)historical panel paintings (Germanisches Nationalmuseum, Cranach portraits), VIS, IRR, UV, X-ray and macro photographs, same painting225 test pairs: VIS-XR 58, XR-IRR 18, XR-MACRO 68, VIS-IRR 60, VIS-UV 21; earlier 2021 set 13 pairs per domainmanual control point pairs on craquelure, mean 122 per pair (75-193 by domain), by Sindel, Maier, Christlein (FAU); 40 points per pair in the 2021 setboth imageshiddenimages credited 'all rights reserved' to the museums; no data release statedaskleft out
its images or its truth are not released
Best painting benchmark found: non-rigid, mixed resolution (scale 1.7-6.5), expert points on both images. Not public.Ask FAU/GNM to release the test set; else rebuild on Closer to Van Eyck images with 20 craquelure points per pair by two people
Fibre breaks in composites, synchrotron CT for DVC (Lee 2026)in situ synchrotron CT (650 nm voxels) of a fibre composite under tension around fibre breaksload steps for DVC (count in the attached PDF, not read)noneone imagedirectcc-by-4.0opentruth to make
3D synchrotron CT load steps, no truth; fibre breaks marked in both volumes or a known warp would make it
Fails truth. Fine-scale aligned-fibre texture, strongly anisotropic.Fibre break positions marked in both volumes give sparse landmarks; known warp on the first scan for dense truth.
Fijiyama test data (multimodal plant imaging)plugin test data; the Fijiyama paper registers MRI and X-ray CT of grapevine trunks and time series (contents of the zip unverified)unverifiednonedirectCC BY 4.0opentruth to make
four 3D test cases (MRI, X-ray CT, photo stacks, an MRI time series); its point sets go with the authors' registered outputs, so landmarks by two people are needed
open the zip to confirm the modalitiestwo people mark 10 wood landmarks (knots, pith) on CT and MRI
FlyingChairs, FlyingChairs2, ChairsSDHomsynthetic: rendered chairs on Flickr backgrounds with affine motion22,872 pairs (FlyingChairs); FlyingChairs2 adds occlusion/backward flow; ChairsSDHom small motionsexact dense flow by construction (Dosovitskiy et al., Ilg et al., Freiburg)both imagesdirectThe dataset is provided for research purposes only and without any warranty. Any commercial use is prohibited.non-commercialleft outpriority 3
Not biomedical (rendered chairs on Flickr photos) and its licence forbids commercial use: outside the corpus rule of 2026-10-27 (open, biomedical).
Planar chairs, unrealistic motion: big domain gap; training set more than benchmark.
FM-SEM GFRP/PP composite multimodal microscopy (Data in Brief 2026)fluorescence, polarized light and backscattered SEM images of the same polished composite cross-sections3 compaction ratios, millimetre-scale stitched images, 15+ bundles eachauthors rigidly registered FM/PLM/SEM stacks; segmentation masksone imagedirectunverified (Recherche Data Gouv)unknowntruth to make
FM, polarised-light and SEM images of the same sections registered rigidly by the authors: fibre centroids matched by a program and checked by a person would make it; licence unverified
Paper calls it a benchmark for registration. Residual SEM brightness drift and stitching artefacts noted.fibre centroids detected in FM and SEM matched by a program, a person verifies 50 per image
Freiburg Scene Flow datasets (FlyingThings3D, Monkaa, Driving)synthetic Blender renders: random flying objects, an animated short film, a driving scene~39k stereo frames in total; FlyingThings3D test split ~4k (paper)exact optical flow, disparity and disparity change by rendering (Mayer et al., Freiburg)both imagesdirectThis dataset is provided for research purposes only and without any warranty. Any commercial use is prohibited.non-commercialto buildpriority 3
Exact rendered flow, research use only, tens of GB a pass; a training set of every flow network
No commercial licence possible (assets); sporadic NaN pixels; no metric scale.
FVC2002 DB1 with manual minutiae (Kayaoglu et al. 2013)fingerprints, 8 impressions per finger with different placement and pressure, same finger100 fingers x 8 impressions in DB1_A (28 pairs per finger)minutiae marked by people on each image (TUBITAK-BILGEM); correspondences between impressions not statedboth imagesrequestunverified (FVC images from Univ. Bologna BioLab; minutiae templates promised in ISO 19794-2, hosting unverified)unknownneeds a request
the data needs a request or a data use agreement
Points are per image, not paired; pairing needs a matching step.Examiners mark 12 corresponding minutiae on 50 impression pairs; or pair existing minutiae after an expert-verified alignment
Gildehauser sandstone imbibition time series (Digital Rocks DRP 134)Fast synchrotron micro-CT of one rock core: dry scan plus a time sequence every 40 s during imbibition, segmented into rock, oil, water; same object over timeDry scan to each time frame; frame to frameNone for motion; segmentations by the authorsone imagedirectODC-BY 1.0opentruth to make
3D synchrotron micro-CT imbibition series, no motion truth; identity on the segmented rock plus a known warp would make it
Rock is rigid: true motion is identity or small sample drift; fluid changes act as appearance change. Other DRP 4D sets (e.g. DRP-1131 Bentheimer) similar, licence unverified.Identity truth for the rock phase (mask from segmentation) plus a known smooth warp to make a non-rigid test; checks robustness to large appearance change
Girl with a Pearl Earring technical examination (Mauritshuis)one Vermeer painting, MA-XRF, IRR, near-infrared, macro and 3D microscopy, same paintingone painting, several modalitiesaligned by the research team for comparison; no points or stack releasedrequestpress kit: images may not be used for commercial purposesaskneeds a request
the data needs a request or a data use agreement
Single object, viewer only; not a dataset as published.Ask Mauritshuis for the registered stack; else click 30 craquelure points on VIS vs IRR
Glass fibre-epoxy compressive failure, ultra-fast synchrotron CT (Wang & Withers)notched unidirectional GFRP under compression (SLS TOMCAT), initial scan, 200 N, 600 N, then streaming to kink-band failureinitial + 2 static load steps + continuous seriesnoneone imagedirectcc-by-nc-sa-4.0non-commercialtruth to make
3D synchrotron CT load steps (about 11 GB per scan), no truth; fibre breaks or pores tracked by hand would make it
Composites in situ load (in scope); fibres give strong texture.Static steps: DVC field cross-checked by tracking 20 fibre breaks or pores by hand; initial vs 200 N near-rigid
HADES triaxial time-lapse microtomography series (Renard et al., Norstore)monzonite, Leitha limestone and Fontainebleau sandstone loaded to failure in the HADES rig at ESRF ID19, time-lapse synchrotron uCTtime series per sample (counts not stated on the page); several samples and recordsnone; DVC strain fields in the papersone imagedirectCC-BY-4.0opentruth to make
3D synchrotron microtomography time series (458.6 GB monzonite record), no truth; grains hand-tracked on low-stress steps would make it
Other records: 10.11582/2019.00015 (Leitha), 10.11582/2018.00022 (Fontainebleau), 10.11582/2024.00029 (filtered). Very large volumes.Low-stress steps near-rigid; DVC verified by hand-tracking 20 grains per pair
HD1K (HCI Benchmark Suite)real urban driving stereo/flow, 2560x1080, varied light and weather~1,000 training frames with flow; half of ground truth withheld for testingflow from lidar static scene + camera pose with per-pixel uncertainty; dynamic objects masked (Kondermann et al., HCI + Bosch)both imagesdirectunverified (benchmark host hci-benchmark.org unreachable 2026-10-10)unknownleft out
The official hosts did not answer on 2026-10-10 nor on 2026-10-12 (HTTP 500, no connection); its truth is lidar plus pose on static parts, half withheld
Accurate truth only on static regions; host site down when checked.
Heritage-Recon (Internet photos aligned to heritage LiDAR scans)Internet photo collections (2D) against terrestrial LiDAR scans (3D) of 4 landmarks (Brandenburg Gate, Pantheon, Lincoln Memorial, Palacio de Bellas Artes)4 scenes, hundreds to thousands of photos eachSfM camera poses with the LiDAR scan (Open Heritage 3D) aligned to the SfM cloudboth imagesdirectcode NOASSERTION; scans from Open Heritage 3D under their own licences (unverified)unknownwaits for a tool
Internet photos against heritage LiDAR scans: waits for a 2D/3D pose error
real photos; poses are SfM, LiDAR alignment by the authors; a rare photo-to-heritage-scan pairingtwo people click 15 photo-to-scan correspondences on 20 photos per scene to bound the SfM-LiDAR alignment
Hilti SLAM Challenge datasets (2022, 2023)multi-camera images and LiDAR (2D and 3D) on construction sites against terrestrial-scan or total-station referencesabout 10-20 sequences per year, construction sites and buildingssparse ground-truth points from total station or TLS-registered control points (sub-cm)both imagesdirectnot read (unverified)unknownwaits for a tool
camera images and LiDAR point clouds against surveyed control points: waits for a 2D/3D pose error and mesh scoring
real; truth is sparse control points, not full poses; dataset terms not read
HisMap multi-temporal historical topographic maps (Zenodo 21707982)scanned German topographic map sheets of the same areas from 1898 to 2023map editions of Hameln and Bavarialand-cover masks; sheet georeferencing unverifieddirectCC BY 4.0opentruth to make
Editions of the same map sheets 1898-2023 with land-cover masks only (CC BY 4.0, 2.6 GB): grid crossings and road junctions marked by two people would make the truth
style changes between editions are largegrid crossings and road junctions marked on 10 edition pairs by two people
HistImage2021 (historical photographs of Dresden landmarks)historical photographs (2D) with SfM reconstructions (3D) of 4 landmarks near Dresden; pose estimation of archival images96 historical images, 4 scenesmanual and automatic feature matches; COLMAP poses and reconstructionsboth imagesdirect for the tie points (OPARA, no login); images fetched one by one from deutschefotothek.de by their permalinksCC BY-SA 4.0 (OPARA deposit); the images are SLUB / Deutsche Fotothek's, with a no-social-network clause in the READMEaskleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): not biomedical; its licence is not shown to allow commercial use (CC BY-SA 4.0 (OPARA deposit); the images are SLUB / Deutsche Fotothek's, with a no-social-network clause in the README). It was proposed as: The OPARA deposit (CC BY-SA 4.0, a 275 KB zip and a README) holds tie points interactively placed in Metashape on 96 historical photographs (at least 15 an image, about 1.5 px) and URLs, not the images. The images must be fetched one by one from SLUB / Deutsche Fotothek, whose material 'is not allowed to be shared in social networks' (README), a custom term. Four landmarks only: test and development scenes cannot both reach ten.
real archival photos; no laser scan, truth is a COLMAP fitregister a modern TLS or photogrammetric scan of the same buildings and re-estimate historical poses against it with manual points
HOT3D (egocentric hand and object tracking, BOP-H3)Aria and Quest 3 headset video (2D) against scanned object meshes and hand models (3D)19 subjects, 33 objects; 4,117 clips of 150 frames in BOPobject, hand and camera poses from an optical marker motion-capture systemboth imagesaccountMeta HOT3D licence agreementnon-commercialneeds an account
the download needs an account
real; mocap truth is independent of images; test annotations partly hidden
HST Hubble Advanced Products: Single Visit Mosaics aligned to Gaiasky fields, HST ACS/WFC3/WFPC2 exposures, same field across filters, visits and instrumentsall public visits since 1990; fields observed many times give multi-epoch and cross-filter pairsper-exposure WCS fitted to Gaia stars by the STScI pipeline (WCSNAME contains FIT_SVM_GAIA); about 10 mas typical, worse far from the Gaia epochboth imagesdirect: MAST portal or astroquery, public (non exclusive-access) data, per-file downloads, no accountMAST Data Use Policy: most MAST data are in the public domain and have no restrictions on use; acknowledgement of the mission and STScI expected (community HLSPs are CC BY 4.0)openleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): not biomedical. It was proposed as: MAST's HAP single-visit products drizzle every exposure of a visit onto one north-up grid, so pairs within a visit are already aligned; a registration pair has to come from two visits of the same field, each WCS fitted to Gaia by the pipeline when possible (WCSNAME containing FIT_SVM_GAIA), a pipeline fit and not a truth made by people. MAST's data use policy says most hosted data are public domain with no restriction on use (acknowledgement expected); the corpus would place Gaia stars on both images as control points and rank this 2D star-field set apart.
MAST says about 80% of ACS/WFC and 50% of WFC3/IR frames were directly aligned; star proper motion adds error between epochs.Gaia DR3 stars propagated to each epoch give exact control points on both images; hold out half for TRE
Huge Strain Challenge (iDICs)2D DIC images at very large strain (content unverified)unverifieddirectnot statedunknownneeds a request
The Drive folder answers 401: the organisers must grant access; content and truth unseen
Current challenge led by J.-F. Witz; only the Drive link was seen. Large-strain decorrelation is the target.Synthetic large-strain warps of a real speckle image (exact), plus painted grid markers on a real specimen.
Image Matching Challenge on Kaggle (2022-2025)photo sets: heritage, day/night, aerial, foliage, transparent objects; 2022 was pairwise fundamental matrixpublic train scenes; test hidden (code competition, ~10k pairs in 2022)camera poses from SfM by the organisers (CTU, UBC, Google, FBK and others)both imagesaccountper-image source and licence in train LICENSE.txt files; competition rules (Kaggle page not readable, unverified)askneeds an account
the download needs an account
Kaggle account needed; test never released; pose truth only.Add hand points or MVS depth on public train scenes for pixel-level error.
Image Matching Challenge PhotoTourism (IMC 2020/2021)tourist photos of landmarks (YFCC100M), wide baseline, different cameras and years15 scenes with full truth for training (100s-1000s images each); test scenes hiddencamera poses and depth from COLMAP SfM/MVS on the full collections (UBC, CTU, Google organisers)both imagesdirectPhototourism images from YFCC100M (Creative Commons per image); Google Urban track: 'released with a restricted license, and must be deleted by the end of the challenge'askin the corpuspriority 3
In the corpus, ranked apart: one stereo pair from each of the IMC 2021 PhotoTourism test scenes (9, ground truth now released) and validation scenes (3, development), points carried by the organisers' poses and cleaned MVS depth (automatic); no licence stated (ask). MegaDepth stays out.
Pose pseudo truth from SfM, not independent; test hidden on evaluation server. status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Image matching survey 2025).Depth+pose give dense correspondence where MVS depth is valid; add hand points on 20 pairs to measure SfM error.
Indoor-6phone images across days and lighting (2D) against COLMAP point clouds (3D) of 6 indoor scenes6 scenes, multi-session, train/test splitsCOLMAP pseudo ground-truth poses and point cloudsboth imagesdirectCC BY 4.0 (repository licence)openwaits for a tool
phone images against COLMAP point clouds: waits for a 2D/3D pose error
real; open licence is rare for relocalisation datacheck pGT with a second reconstruction tool or with manual 2D-3D clicks on test images
InLociPhone photos (2D) against a large indoor laser-scanned 3D map (WUSTL RGB-D scans)356 query photos, about 10,000 database cutouts with depth6-DoF query poses estimated then manually verified against the 3D scansboth imageshiddennot stated on page (unverified)unknownleft out
its images or its truth are not released
real; query poses kept for the online evaluation tool; database point clouds from laser scans status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Image matching survey 2025).
International PIV Challenge (1st-4th)2D, stereo and 3D (tomographic) PIV images of fluid flows, synthetic and experimental4 editions (2001, 2003, 2005, 2014), 3-6 cases eachknown flow for synthetic cases (2001 B and E, 2003 B, 2005 A and B, 2014 C and D); analytic solid-body rotation for 2014 Case F (real); none for other experimental casesboth imagesdirectnot stated ('Please freely link this web-site')unknownneeds a requestpriority 3
the known flow of the synthetic cases is given on e-mail request only
Whether ground truth fields are downloadable for every synthetic case is unverified.Solid-body rotation and uniform shift cases give analytic truth on real images; extend with known rotations on a turntable.
Inv3D (template-guided invoice dewarping)synthetic invoices rendered on real paper shapes + real photos (Inv3DReal) with their templatessynthetic set + Inv3DReal real photos (counts unverified)synthetic: exact rectification meshes and per-pixel maps; real: flat template only (Hertlein et al., FZI/KIT)both imagesdirectCC BY-NC-SA 4.0 ausgenommen 'restricted-license-files.txt'. Diese sind nur zur akademischen Nutzung.non-commercialleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): not biomedical; its licence is not shown to allow commercial use (CC BY-NC-SA 4.0 ausgenommen 'restricted-license-files.txt'. Diese sind nur zur akademischen Nutzung.). It was proposed as: Exact rectification maps on rendered invoices; the real photos have no correspondences; CC BY-NC-SA 4.0 with academic-only files
A form/template alignment set; real part has no correspondences.Hand points between photo and template on Inv3DReal, or printed markers.
Ionization chamber CAD-to-CT registration (2026)cylindrical ionization chambers, cone-beam CT vs CAD20 simulated + 18 real scansexact pose for simulated scans; real scans checked against CMM interior volumeshiddenno data release statedunknownleft out
its images or its truth are not released
Not released.Their simulation route (CAD to simulated CT) gives exact poses; CMM features for real scans
IPD (Intrinsic Industrial Plenoptic Dataset, BOP-Industrial)multi-view high-resolution HDR photos and structured-light depth (2D) of metal industrial parts against CAD (3D)industrial parts in bins, multi-camera rigs (count not read)6D poses from multi-view annotation with calibrated rig (Kalra et al. CVPR 2024)both imagesdirectCC BY-SA 4.0 (custom licence on request)openwaits for a tool
photos and depth of metal parts against CAD: waits for a 2D/3D pose error
real; reflective metal parts
IPK multimodal plant image registration (mPIR, FLU/VIS/NIR)greenhouse plants on LemnaTec systems, fluorescence vs visible vs near-infrared cameras, same plant same timeunknownno public truth found; how Henke et al. 2018/2019 scored accuracy is unverifiedhiddentool free to download; image data licence unverifiedunknownleft out
its images or its truth are not released
Tool is public, image pairs were not found public. Different cameras and viewpoints: affine plus leaf motion.Ask IPK for pairs; plant mask from FLU (easy) vs VIS gives overlap truth; two people click 10 leaf tips on 20 pairs
ITTO (Is This Tracker On? point tracking benchmark)real videos (existing datasets and egocentric recordings) with long occlusions and complex motioncount unverifiedhuman point annotations from a multi-stage pipelineboth imagesdirectunverifiedunknownwaits for a tool
real video with annotated point tracks: waits for point tracking
Caltech 2025.
KAIST Multispectral Pedestrian (RGB-thermal)colour and thermal video through a beam splitter (aligned by hardware), traffic scenes day and night~95,000 colour-thermal pairspedestrian boxes on the aligned pairboth imagesdirectunverified (GitHub licence: Other)unknowntruth to make
Colour-thermal pairs aligned by a beam splitter: a known warp on the aligned pairs would give an exact two-modality truth; licence not asserted on GitHub
Beam-splitter capture aligns the two modalities by construction (small residual misalignment reported in later work, unverified).warp the thermal frame by a known smooth field and register colour to it: exact 2D multimodal truth
KITTI odometry (camera + Velodyne, used for image-to-point-cloud registration)camera images (2D) against Velodyne LiDAR scans (3D) of driving scenes; I2P registration and camera-LiDAR calibration22 sequences; 11 with poses; I2P papers use seq 00-08 train, 09-10 test with random perturbationsfactory camera-LiDAR extrinsic calibration (target-based) plus GPS/IMU poses; I2P truth is that calibration with a known synthetic perturbationboth imagesaccountCC BY-NC-SA 3.0 (KITTI)non-commercialneeds an account
the download needs an account
real; the truth is a calibration, so it is exact only up to calibration error (about 1 cm, 0.1 deg reported)
KITTI-360perspective and fisheye images (2D) against Velodyne and pushbroom laser scans (3D), suburban Karlsruheabout 320,000 images and 100,000 laser scans over 73.7 kmcamera-LiDAR calibration and GPS/IMU poses refined by large-scale optimisationboth imagesaccountCC BY-NC-SA 3.0non-commercialneeds an account
the download needs an account
real; full 360-degree coverage, accumulated laser maps available
Kubric MOVi (A-F)synthetic multi-object videos (Blender + PyBullet), scanned objects on HDRI backgroundsMOVi-A..F TFDS sets of 24-frame videos; MOVi-F for flow; generator makes unlimited dataexact forward/backward flow, depth and point tracks by rendering (Google Research)both imagesdirectrepository Apache-2.0; no separate data licence seenopenwaits for a tool
synthetic video with exact flow and point tracks: waits for point tracking
256-512 px; the generator can make pairs with known deformation on demand.
Kust4K (pixel-aligned UAV RGB and thermal pairs)UAV RGB and thermal infrared images of urban roads, day and night4,024 pairspairs pixel-aligned by the authors; semantic labelsboth imagesdirectunverified (figshare)unknowntruth to make
UAV RGB-thermal pairs pixel-aligned by the authors: a known warp would make the truth; licence unverified
Sci Data 2025.known warps on the aligned pairs for exact RGB-TIR truth
LaMARAR device images (HoloLens, phone) to laser scans and maps of 3 large scenes (CAB, HGE, LIN); localisation and mapping3 scenes, multi-session trajectories, thousands of queriesposes from automatic alignment of device trajectories against NavVis laser scans (pipeline in paper)both imagesrequestterms through the data access process (unverified)unknownneeds a request
the data needs a request or a data use agreement
real; lidar point clouds and raw sensor data released Oct 2023; access process not checked
LLNL D4DCT dynamic CTsimulated dynamic CT of objects deforming under compression (MPM)2 datasets, 256^3known simulated deformation over timeboth imagesdirectCC BY 4.0opentruth to make
simulated 3D CT frames (256^3, 60 per sample) with projections, but the MPM deformation fields are not in the release; a known warp would make it
Made for reconstruction; deformation fields availability unverified.
LLVIP (visible-infrared paired, low light)visible and thermal infrared images strictly aligned in time and space, mostly dark scenes15,488 pairspedestrian boxesboth imagesrequestfree for non-commercial use (terms on page)non-commercialneeds a request
the data needs a request or a data use agreement
Alignment done by the authors (registration), not by optics; low light makes the visible image very weak.warp the infrared image by a known field and register visible to it: truth exact up to the authors' alignment
M3OT (two-drone RGB and thermal tracking)RGB and thermal video from two UAVs at 100-120 m10,790 RGB-IR image pairsvehicle boxes and track IDs in both modalitiesboth imagesdirectunverified (figshare)unknowntruth to make
RGB and thermal drone video with vehicle boxes and track IDs in both: box centres of one track, checked by a person, would give point pairs; pair alignment unstated; licence unverified
Sci Data 2025. Pair alignment not stated.box centres of the same track ID give point pairs; verify 20 per sequence
M4-SAR optical-SAR object detection pairs (Kaggle mirror)co-registered optical and SAR satellite images, same scenelarge-scale (count unverified)boxes on aligned pairsaccountCC0 on the Kaggle mirror (original licence unverified)unknownneeds an account
the download needs an account
detection set; alignment by the authorsaligned by construction: apply known warps for exact optical-SAR truth
McPed (ground-to-air cross-view pose, field robots)ground RGB, 144-channel LiDAR and aerial orthophotos, fields and urban roads29,940 frames in 11 environmentscentimetre georeferencing from GNSS fusion and PPKboth imagesdirectunverified (figshare)unknownwaits for a tool
ground photos and LiDAR against aerial orthophotos (cross-view pose): waits for a 2D/3D pose error
Sci Data 2025.
MegaDepthinternet photos of ~200 landmarks, pairs across viewpoint, camera and years~130k images; standard MegaDepth-1500 test pairs (LoFTR protocol)COLMAP poses + MVS depth (Li and Snavely, Cornell); dense correspondence by depth reprojectionboth imagesdirect (cs.cornell.edu), no account, but only as MegaDepth_v1.tar.gz (199 GB) and MegaDepth_SfM_v1.tar.xz (667 GB)depth maps and SfM models CC BY 4.0; original Flickr images under their own, unstated licencesaskleft out
The correspondences would come from COLMAP SfM poses and cleaned MVS depth (automatic pseudo truth), and the project page licenses the depth maps and SfM models under CC BY 4.0 while 'the original images come with their own licenses', none stated. The images and depths are one 199 GB tar.gz and the poses are in a separate 667 GB tar.xz of SfM models, neither readable by ranges; wide-baseline photographs are already in the corpus (HPatches, WxBS).
Pseudo truth: depth holes and errors at edges, transient objects. status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Image matching survey 2025).Keep pixels with consistent depth; hand points on 20 pairs to estimate truth error.
MIAS-LCEC datasets (LiDAR-camera extrinsic calibration)RGB images (2D) against Livox solid-state LiDAR point clouds (3D), indoor and outdoor, target-free and checkerboardTF70: 60 pairs; CB70: 15 checkerboard pairs; TF360: 12 pairsreference extrinsics (method per subset not read; checkerboard subset likely target-based)both imagesdirectnot seen on page (unverified)unknownwaits for a tool
RGB images against LiDAR point clouds (extrinsics): waits for a 2D/3D pose error
real; small; a benchmark of the calibration problem itselfuse the checkerboard subset as truth and score target-free methods on the same rig
Middlebury optical flowsmall real and synthetic scenes; real truth via hidden fluorescent texture8 public pairs with truth (RubberWhale, Hydrangea, Dimetrodon, Grove2/3, Urban2/3, Venus); 8 eval pairs hiddendense flow from UV-lit hidden texture (real), rendering (synthetic), stereo (Baker et al., Middlebury)both imagesdirectnone stated on the data pageaskin the corpuspriority 3
In the corpus, ranked apart: the eight scenes with public flow (3 hidden-texture, 4 rendered, 1 stereo), all test, 245 to 256 points a pair from the flow; no licence stated (ask).
Tiny, small motions, saturated; real truth is high quality.
MNIST (digit images for registration toys)28x28 digit images, between samples of the same digitthousandsnonedirectnot stated on the page (CC BY-SA 3.0 often cited, unverified)unknownleft out
28 px digit glyphs with no truth: a toy with nothing an imaging benchmark needs; the corpus's known warps belong on real images
Toy only; no correspondence truth.known random warps applied to digits give exact truth
MRSR (RGB-SAR and RGB-IR remote sensing registration set)optical to SAR and optical to infrared satellite patches3,850 RGB-SAR + 4,000 RGB-IR pairspre-aligned pairs with synthetic transforms (unverified)directunknownunknownleft out
No download found: the set is only named in a 2026 paper
Details unverified; name as cited in the paper.known affine or smooth warps on aligned pairs give exact truth
MV-TAP evaluation sets (multi-view point tracking)multi-view videos of dynamic scenes with point trackssynthetic training set plus real evaluation sets (counts unverified)tracks per view (method for real sets unverified)hiddenunverifiedunknownleft out
its images or its truth are not released
Release not confirmed.human verification of 20 tracks per real sequence
MVSEC (event optical flow)event + grayscale stereo from car, drone, motorbike, handheld; indoor and outdoorindoor_flying and outdoor_day/night sequences with computed flowflow from lidar depth + pose (mocap indoors, odometry outdoors) (Zhu et al., UPenn)both imagesdirect: ROS bags on visiondata.cis.upenn.edu, HDF5 files and ground-truth flow npz in public Google Drive folders, no accountCreative Commons Attribution-ShareAlike 4.0 International (project page)openleft outpriority 3
At most 6 independent scenes (the four indoor flights share one room, the five drives are in West Philadelphia), with flow on the frames' own grid for at most 3 (the night sequences' flow is given undistorted only); the event streams, the set's point, are not images the corpus can use. Automatic truth; CC BY-SA 4.0.
Low resolution (346x260); static-scene assumption.
MVTec ITODD and ITODD-MV (BOP-Industrial)grey and structured-light depth images (2D) of 28 industrial objects against CAD models (3D); single and multi-view28 objects; ITODD-MV adds 3 high-resolution grey views per scene6D poses annotated by MVTec; test poses hidden (validation public)both imageshiddenCC BY-NC-SA 4.0 (custom licence on request)non-commercialleft out
its images or its truth are not released
real; untextured CAD only, like X-ray-to-CAD inspection
NAVER LABS large-scale indoor localization (Gangnam Station, Hyundai Department Store)phone and camera images in crowded indoor spaces (2D) against lidar-based maps (3D); 5 datasetsabout 130,000 images over 5 sites; mapping, validation and test splits6-DoF poses from a mapping robot with lidar SLAM and SfM refinement; test poses withheldboth imagesrequestBY-NC-ND with a custom research-derivatives clause; full text in license.txtnon-commercialneeds a request
the data needs a request or a data use agreement
real; kapture format; sparse lidar depth for training images; access by request through NAVER LABS open dataset hub
Newer College datasethandheld stereo/multi-camera images and LiDAR (2D and 3D) against a survey-grade prior laser map (3D) of New College, Oxford2020 stereo set (about 2.2 km) and 2021 multi-camera extension6-DoF device poses from ICP of each LiDAR scan to the 290-million-point prior map (about 3 cm)both imagesdirectnot read (project page; unverified)unknownwaits for a tool
camera images and LiDAR scans against a prior laser map: waits for a 2D/3D pose error and mesh scoring
real; image-to-map truth via LiDAR-camera extrinsics; URL is the paper
Niantic Map-free Relocalizationone reference photo plus one query photo per scene, metric relative pose (2D-to-2D with scale); 655 outdoor places655 outdoor places: 460 train, 65 validation, 130 test, each a reference frame plus query sequenceposes from SfM over full scans with metric scale from phone AR tracking (test poses hidden)both imageshiddenNiantic Map Free Relocalization Dataset License: non-commercial onlynon-commercialleft out
its images or its truth are not released
real; strictly no 3D map at query time, so it is 2D-2D metric pose, not 2D-3D; listed for completeness
Nickel superalloy cryogenic deformation, HR-DIC with EBSD (Dryad 2024)SEM high-resolution DIC images during loading down to liquid helium, with EBSD maps of the same areaload steps (count unverified)none for registration; DIC fields are method outputsdirectCC0opentruth to make
SEM HR-DIC images with EBSD maps and no registration truth (CC0, 9.1 GB): grain triple points from EBSD found on the SEM images and checked by a person would make it
grain boundary triple points from EBSD as landmarks in the SEM images, checked by a person
NIST AMMT overhang parts: XCT + design STLlaser powder bed fusion metal overhang parts, XCT vs as-designed STL (CAD nominal)4 parts (X4); a sister set of 16 parts (mds2-2309)design STL as nominal; XCT aligned to part datums (3-2-1) in VGStudio by NIST; no XCT-to-CAD alignment givenone imagedirecthttps://www.nist.gov/director/licensing (as listed)openwaits for a tool
XCT volumes against as-designed STL meshes, no XCT-to-CAD alignment given: waits for mesh scoring
Nominal-actual: overhang droop is the deviation to measure, so only the rigid part is truth.Datum-plane (3-2-1) alignment as reference; add CMM measurements for deviation truth
NIST Special Database 302 (annotated latent and exemplar fingerprints)fingerprints, latent vs exemplar impressions of the same finger, 2D non-rigid10,000 latent images with exemplars; SD 302i holds latent-exemplar minutiae correspondencesminutiae, orientation and ridge quality marked by Certified Latent Print Examiners (NIST); correspondences in SD 302iboth imagesrequest'freely available for research purposes' (no licence named)askneeds a request
the data needs a request or a data use agreement
Biometric data reviewed by NIST's Research Protection Office; skin distortion gives a real 2D non-rigid problem with expert point pairs.
NoW benchmark (single image to 3D face scan)photos of faces (2D) against high-resolution 3D head scans of the same 100 subjects (3D); reconstruction scored after rigid scan alignment2,054 images of 100 subjects (validation scans public, test scans hidden)active-stereo 3D scans; evaluation by landmark-initialised rigid alignment and scan-to-mesh distanceboth imageshiddenMPG licence after registration (non-commercial research; unverified)non-commercialleft out
its images or its truth are not released
real; a reconstruction rather than registration benchmark, but its truth is a 2D-to-3D scan correspondence
nuScenes (camera + LiDAR)6 cameras (2D) against a 32-beam LiDAR (3D), Boston and Singapore driving1000 scenes (850 train/val, 150 test), 40,000 keyframescalibrated camera-LiDAR extrinsics; I2P truth is calibration plus synthetic perturbationboth imagesaccountCC BY-NC-SA 4.0 (non-commercial; commercial licence on request)non-commercialneeds an account
the download needs an account
real; sparse LiDAR makes 2D-3D matching harder than KITTIDerive piecewise-rigid flow from tracked boxes and ego pose (AV2 recipe), then hand-check a sample.
Open Heritage 3D (CyArk and partners: heritage LiDAR scans with photogrammetry imagery)terrestrial and aerial photographs (2D) with terrestrial/aerial LiDAR scans (3D) of heritage sites worldwidehundreds of sites, about 25 GB per site; raw images and raw scans for manynone packaged for registration; providers co-register scans and photogrammetry (often with surveyed control points, per project)one imagerequestper project, several Creative Commons licences (e.g. CC BY-NC 4.0); download links sent after a name/email formnon-commercialneeds a request
the data needs a request or a data use agreement
real; raw material for a photo-to-scan benchmark rather than a benchmark; commercial use excluded on many sitespick 5 sites with raw photos and TLS: run SfM on photos, align to TLS with surveyed targets where present, then two people click 15 photo-to-scan points per site to validate
OpenCVL (fine-grained cross-view localisation)ground-level images aligned to geo-referenced aerial imagery, 41 European cities617,388 ground-aerial pairspose annotations, curated and corrected for in-the-wild imagesboth imagesdirectunverified (permissive sources)unknownwaits for a tool
ground photos against aerial images (cross-view pose): waits for a 2D/3D pose error
Earlier street-to-satellite sets were listed as failing the pixel criterion; this one gives fine pose.
Oxford Affine (Mikolajczyk) sequencesreal photos of planar or distant scenes: viewpoint, zoom/rotation, blur, light, JPEG8 sequences x 5 pairs = 40 pairshomography per pair, by the authors (Mikolajczyk et al., 2005)both imagesdirectnone stated on the pageunknownleft out
Its viewpoint and zoom sequences are in HPatches (v_graffiti, v_bricks, v_bark, v_boat, i_leuven); the other three are blur and JPEG series from a fixed camera with nothing to register
Tiny and old; superseded by HPatches but still cited.Re-check homographies with hand-placed points.
Oxford Day-and-Night (egocentric relocalisation under lighting change)Meta Aria egocentric video of the same places by day and night30 km of trajectories over 40,000 m2multi-session SLAM poses and 3D point clouds aligned across sessionsboth imagesdirectunverifiedunknownwaits for a tool
egocentric video against SLAM point clouds (relocalisation): waits for a 2D/3D pose error
Oxford RobotCarcar cameras (2D) against 2D/3D LiDAR scans (3D) over 100+ repetitions of one Oxford routeover 100 traversals across a year; RTK ground truth released 2020RTK-GNSS/INS poses and calibrated extrinsics; used to build 2D3D-MatchNet and RobotCar Seasonsboth imagesaccountCC BY-NC-SA 4.0 (registration required; unverified)non-commercialneeds an account
the download needs an account
real
Oxford Spirescolour camera images and mobile LiDAR (2D and 3D) against millimetre-accurate terrestrial laser scans (3D) of Oxford landmarks24 sequences over 6 sites (about 1 ha each)sensor trajectory from registering mobile LiDAR to the Leica RTC360 TLS map (cm-level); calibrated camera-LiDAR rigboth imagesdirectCC BY-NC-SA 4.0 (Hugging Face card)non-commercialwaits for a tool
camera images and mobile LiDAR against terrestrial laser scans: waits for a 2D/3D pose error and mesh scoring
real; camera poses follow from LiDAR-to-TLS registration plus extrinsics, so they are independent of image content
PASCAL3D+PASCAL VOC and ImageNet photos (2D) aligned to approximate CAD models (3D) of 12 rigid categoriesover 30,000 annotated instancesviewpoint from annotators aligning one of a few CAD models per category with clicked keypoints (approximate shape)both imagesdirectnot stated (images under PASCAL/ImageNet terms; unverified)unknownwaits for a tool
photos against approximate CAD models: waits for a 2D/3D pose error
real; CAD is only category-level approximate, so pose truth is coarse
Pheno4D (maize and tomato 4D point clouds)maize and tomato plants, laser-scanned 3D point clouds, same plant over 12-20 days7 maize x 12 days, 7 tomato x 20 days; 49 + 77 labelled cloudshand-labelled organs, leaf IDs consistent over time; clouds registered in a common frame (Schunck et al. 2021, Univ. Bonn)both imagesdirectNo licence on the page; Hugging Face mirror says contact the maintainer (L. Klingbeil) before redistributionaskwaits for a tool
laser-scanned point clouds of growing plants, no images: waits for mesh scoring
Point clouds, not images; growth makes it non-rigid. Leaf IDs give correspondences.Leaf tip and collar points per leaf ID across days as landmarks (derivable from the labels)
Pix3DInternet photos of IKEA furniture (2D) aligned to exact CAD models (3D); single-image pose and shapeabout 10,000 image-shape pairs, 395 shapes, 9 categoriespose fitted from manually clicked 2D-3D keypoints plus PnP and refinement (Sun et al. 2018)both imagesdirectnot seen on page (unverified)unknownwaits for a tool
photos against exact CAD models: waits for a 2D/3D pose error
real photos with exact CAD; truth is human keypoints, so pose error of a few degreessecond annotator re-clicks keypoints on 200 images to measure pose uncertainty
Plant Phenotyping Datasets (CVPPP, Minervini/Scharr/Tsaftaris)Arabidopsis and tobacco rosettes, top-view RGB time-lapse, same plant over daystrays imaged over days; about 800 labelled images in the LSC subsets A1-A4expert leaf instance masks; page lists leaf tracking among tasks (whether leaf IDs stay consistent over time is unverified)both imagesaccount"Data cannot be used for commercial use ... Data can be used for academic purposes in the context of developing and evaluating algorithms." Also: no redistribution, no use in challenges without owners' consentnon-commercialneeds an account
the download needs an account
Terms forbid redistribution and challenge use without consent: a benchmark needs the owners' permission. Leaf growth and new leaves make it non-diffeomorphic.Leaf masks with consistent IDs on consecutive days as overlap truth; two annotators click leaf tips on 10 pairs
PointOdysseysynthetic long videos of animated humans and animals in varied scenes159 videos (131 train, 15 val, 13 test), v1.2exact long-range point tracks, depth and masks by rendering (Zheng, Harley et al., Stanford)both imagesdirectHugging Face card (aharley/pointodyssey): MIT; site footer: 'Copyright 2023. All rights reserved.'openwaits for a tool
synthetic long video with exact point tracks: waits for point tracking
Synthetic; licence wording conflicts between site and HF card.
Polycrystalline metal: 3D microstructure and deformation fields (Dryad 2022)spatially correlated 3D microstructure and surface deformation fields of the same specimen1 specimen, several load steps (unverified)authors' spatial correlation between modalitiesdirectCC0opentruth to make
Correlated microstructure and surface strain maps of one specimen, related by the authors (CC0, 55 GB): grain triple points checked by a person would make a truth, on one specimen only
grain triple points as landmarks across modalities
PragueParks (Image Matching Challenge 2021)video frames of park scenes, two viewsunverifiedposes, test hidden on serverboth imageshiddenunverifiedunknownleft out
its images or its truth are not released
IMC 2021 stereo track. Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Image matching survey 2025).
RaFD (Radboud Faces Database)2D face photos, same person across 8 expressions and 3 gaze directions67 people, 1,608 imagesnone given; papers use dlib landmarks or the neutral face as targetrequestresearch use after approval (unverified, site did not answer)non-commercialneeds a request
the data needs a request or a data use agreement
Non-medical 2D toy for deformable registration.detected facial landmarks (dlib) on both, checked by a person on 20 pairs
RGB-D Scenes v2 (image-to-point-cloud protocol)RGB image (2D) against point cloud fragments fused from other frames (3D) of tabletop scenes14 scenes; 1,748 train / 236 val / 497 test pairs in the 2D3D-MATR protocolcamera poses from RGB-D reconstruction of each sceneboth imagesdirectnot stated on page (unverified)unknownwaits for a tool
RGB images against point-cloud fragments: waits for a 2D/3D pose error
real; archived UW page, scene poses come from Patch Volumes reconstruction; no licence statedkeep reconstruction poses; add manual 2D-3D clicks on the 497 test pairs to bound pGT error
RGB-NIR Scene Dataset (EPFL)visible and near-infrared photos of the same scene, separate exposures, 9 scene categories477 pairsalignment by SIFT + RANSAC similarity transform (authors)both imagesdirectunverifiedunknowntruth to make
Visible and NIR exposures aligned only by a similarity, so near objects keep parallax: points by people on 20 pairs would make a two-modality truth; licence unverified
Aligned by a similarity transform after capture, not by construction: residual parallax for near objects.warp NIR by a known smooth field and register RGB to it: truth exact up to the original similarity alignment
RGBT-3M UAV RGB-thermal wildfire pairs (Kaggle mirror)drone RGB vs thermal images of wildfire scenesregistered pairs (count unverified)authors' registration plus smoke, fire and person labelsaccountOther (specified in description)unknownneeds an account
the download needs an account
mirror of Remote Sensing 2025, 17(15), 2593 (USTC); original on Quark Netdiskknown warps on the registered pairs give exact multimodal truth
Rice root daily X-ray CT series (Teramoto, NARO)upland rice root system in a soil pot, X-ray CT, same pot daily 7-27 days after sowing21 daily volumes of one pot (demo data, 8-bit normalised)rigid alignment by ICP on bright soil particles (algorithmic, Teramoto & Uga 2022); manual root vectors at the last time point in RSAtrace4Done imagedirectNo licence stated for the data; repo licence NOASSERTIONunknowntruth to make
3D CT of one pot over 21 days, aligned only by ICP; soil particles verified by a person would give landmarks
Project site rootomics.dna.affrc.go.jp did not resolve for me; download link taken from the README.Soil particles are dense natural fiducials: a person verifies 20 matched particles per pair; rescan the same pot twice in a day for a zero-motion pair
RIO10RGB frames (2D) against reconstructed meshes (3D) of indoor rooms that change over time (3RScan); long-term relocalisation74 sequences over 10 scenes: 10 train, 10 val, 54 testposes from 3RScan RGB-D reconstruction and scan-to-scan alignment; test poses on evaluation serverboth imagesrequest3RScan terms of use (Google form)unknownneeds a request
the data needs a request or a data use agreement
real; scenes change between map and query (moved furniture), a rare test of change
RobotCar Seasons (v1, v2)car camera images (day, night, rain, snow) against an SfM/LiDAR-based map of Oxford26,121 reference, 11,934 query images (v1); v2 adds training posesquery poses from LiDAR-aided SfM and manual 2D-3D correspondences for night (Sattler et al. 2018)both imageshiddenOxford RobotCar licence is CC BY-NC-SA 4.0 (base data); readme terms unverifiednon-commercialleft out
its images or its truth are not released
real; test poses hidden
RobustSpring (corrupted Spring for flow, scene flow, stereo)Spring rendered scenes with 20 image corruptions applied consistently in time, stereo and depth20,000 corrupted imagesSpring ground-truth flow and disparityboth imagesdirectunverified (Spring terms)unknownleft out
Built on Spring's 2,000 test frames, whose flow truth is withheld; its robustness metric needs no truth
Derived from Spring (listed).
Root-reinforced soil 4D x-ray CT (Bull, Smethurst et al.)x-ray CT of soil with plant roots under direct shear, several root speciesshear series per sample, plus a 4-scan noise study (record 3361832, 1800x1800x1400 uint8)zero motion for the noise study repeats; otherwise author DVC onlyone imagedirectcc-by-4.0opentruth to make
3D CT shear series with author DVC only; the 4-scan noise study gives zero-motion pairs, root branch points traced in both would give landmarks
Sibling Zenodo records (3556741, 3558554 and others) hold the shear-zone series.Root branch points traced in both volumes as landmarks; noise-study repeats as zero-motion truth.
RUBIK (image matching benchmark by geometric difficulty)driving image pairs from nuScenes graded by overlap, scale ratio and viewpoint angle16.5K pairs in 33 difficulty levelsrelative poses from nuScenes sensorsboth imagesdirectunverified (nuScenes licence applies)non-commercialneeds an account
its images are the nuScenes test set, whose download needs an account (2D pairs, then the 2D builder)
2025.
Sandstone hydraulic fracturing in situ CT (Ma, Wang, Zheng)sandstone under triaxial stress with water injection, in situ CT during fracturing, same sampleseveral scans across the test (count unverified)none; pressure logs onlyone imagedirectCreative Commons Attribution 4.0 Internationalopentruth to make
in situ CT scans of one sandstone sample, no truth; 20 points on fracture-free regions of pre-fracture scans would make it
Fracture opening is a discontinuity: hard case for smooth registration.Pre-fracture scans are near-rigid; click 20 points on fracture-free regions
ScanNet (v2)RGB-D frames (2D) against reconstructed meshes (3D) of 1,513 indoor scans; used for image-to-point-cloud matching (P2-Net, 2D3D-MATR variants)1,513 scans of 707 spaces, 2.5 million framescamera poses from BundleFusion reconstruction (pseudo ground truth)both imagesrequestScanNet Terms of Use (signed agreement, non-commercial)non-commercialneeds a request
the data needs a request or a data use agreement
real; truth is a reconstruction; request by signed terms status change: Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Image matching survey 2025).Use ScanNet++ laser geometry instead, or hand points on test pairs.
ScanNet++33 MP DSLR images and iPhone RGB-D (2D) registered to sub-millimetre laser scans (3D) of 1000+ indoor scenes1000+ scenes (v2), each with laser scan, DSLR set and iPhone streamimage poses from COLMAP registered to the laser scan, refined against scan renderingsboth imagesaccountScanNet++ Terms of Use (application and token required)non-commercialneeds an account
the download needs an account
real; application must be approved before download; terms forbid commercial use (unverified detail)
SDO ML dataset (SDOML v1/v2, AIA + HMI)Sun, SDO AIA EUV/UV channels and HMI vector magnetograms, same instant, multichannel and cross-instrument 2D2010-2020 (v2), AIA every 6 min, HMI every 12 min: millions of channel pairsall channels resampled to one co-registered 512x512 grid by the makers (Galvez et al. 2019, NASA FDL), from instrument pointingboth imagesdirectAWS registry (v2): "There are no restrictions on the use of this data."; Stanford v1 yearly records: CC BY 3.0opentruth to make
AIA and HMI channels resampled to one 512x512 grid: a known warp on one channel would make an exact two-modality truth; no restrictions on use
Downsampled from 4096 px; alignment is the pipeline's, not independently checked. Truth is zero transform between channels.Apply known rigid/smooth warps to one channel (exact multimodal truth since channels share a grid); full-res level-1 data plus limb fits for a harder version
SEM-DIC Round Robin (iDICs)scanning electron microscope speckle images for DIC (content unverified)unverified; SEM DIC usually relies on stage translations to calibrate drift and spatial distortiondirectnot statedunknownneeds a request
The Drive folder answers 401: the organisers must grant access; content and truth unseen
Current round robin; only the Drive link was seen. SEM drift and scan-line shifts make it a hard registration case.Known stage translations in two orthogonal directions, as in Kammers and Daly 2013, give rigid truth.
Sentinel-2 Monesi landslide images for DIC (Notti 2021)Sentinel-2 RGB-NIR images of the same landslide, 2016 and 20181 pair plus landslide outlinenone; landslide outline onlyone imagedirectcc-by-4.0opentruth to make
One Sentinel-2 pair with a landslide outline and no truth (CC BY 4.0): stable ground outside the outline as zero motion is all it offers, on one pair
Fails truth. Shared to train DIC on satellite images.Stable ground outside the outline as zero-motion truth; GNSS or field survey points if the authors have them.
ShapeNetCAD models of objectsbetween shapescategoriesaccountresearch termsnon-commercialneeds an account
the download needs an account
Non-medical shape set; point-cloud registration only. Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Point-cloud shape survey 2025).CAD models, not images: render two views with known poses for a 2D/3D toy truth only.
SHREC 2021 cryo-ET (simulated cell-like tomograms)simulated tilt series (2D projections) and tomograms (3D) of cell-like volumes with 12 protein complexes10 volumes (model 0-9)exact particle positions and class masks by construction (orientations included, per memory; unverified)both imagesdirectCC0 1.0 (Dataverse NL doi:10.34894/XRTJMA)opentruth to make
simulated 3D tomograms with their noise-free volumes and class masks, aligned by construction; a known warp would make the pair
synthetic; class 4V94 built mirrored and dropped from evaluation
SID-CV (pan-Arctic sea ice drift from Sentinel-1 pairs)sequential Sentinel-1 SAR image pairs 0-36 h apart, drifting sea ice350k+ products, 2017-2021drift vectors from feature tracking and pattern matching (algorithm output), validated against buoysone imagedirectunverified (Science Data Bank)unknowntruth to make
Its drift vectors are a feature-tracking result, a registration output rather than a truth; buoy positions give sparse independent points; licence unverified
Truth is itself a registration result; buoys give sparse independent checks.use buoy positions as sparse independent landmarks
SMART-Ship (5-modality registered satellite sets of berthed ships)visible, SAR, panchromatic, multispectral and NIR satellite images of the same ports within a week1,092 image sets, 38,838 shipsauthors registered sets; ship instance IDs and polygons on each modalityboth imagesdirectunverifiedunknowntruth to make
Five satellite modalities registered by the authors, with ship polygons: port structures marked by people would be the truth (berthed ships can move between dates); licence unverified
2025.ship polygon centroids give multimodal point pairs; a person checks 20 per set
spam tutorial datasets (Ando, Stamati et al.)x-ray CT of notched sandstone before and after deformation (VEC4); triaxial compression series of a Martian simulant sand (M2EA05); paired x-ray and neutron CT of sand and concreteVEC4 1 pair; M2EA05 a load series; SandNX and YehyaConcreteNX x-ray/neutron pairsnone (author DVC and registration results are not truth)one imagedirectcc-by-4.0opentruth to make
3D x-ray and neutron CT of sand, sandstone and concrete, no truth; tracked grains or a known warp would make it
Includes a true multimodal 3D pair (x-ray vs neutron) outside medicine.Track individual sand grains between scans (spam discrete DVC) and verify a sample by eye; for x-ray/neutron, warp one modality with a known field after aligning.
STAMM3D multimodal rigid 3D benchmarkrigid 3D registration across lab CT, synchrotron CT, MRI and CAD of objects24 archives (pair count unverified)unverified (empty description)Zenodo, direct; RAR archives, not readable by rangesCC BY 4.0 (Zenodo)openleft out
24 RAR archives (14.3 GB) of lab CT at 60 and 130 kV, a commercial CT, MRI, synchrotron CT and CAD of about seven objects (a nut, a leg, a card, 'BBI', 'SBI', a brachiopod, a GC model), with 19 rigid ground-truth matrices and their inverses (file names in gt_matrix_storage.rar); the record's description is one line and no paper says how the matrices were obtained. Rigid by design, under ten objects, and the truth's origin could not be stated in a descriptor.
No paper found; file names suggest nuts, fossils and printed parts.fiducial beads on the objects or CAD geometry as reference
Stationary images for DIC uncertainty (stereo-DIC, Zenodo 13375616)repeated stereo-DIC images of a stationary speckled specimenimage serieszero displacement by designboth imagesZenodo, direct, one 5 GB zipCC BY 4.0 (Zenodo)openleft out
Four sets of 100 stereo images (5 MPx) of one speckled specimen standing still, set 2 after a 1 mm rigid shift, with stereo calibrations (Zenodo record): one specimen, zero or rigid motion, a noise-floor test. A known warp on it would repeat one speckle pattern; the DIC Challenge rows (registration 1427, 1428, 1461) already offer speckle images under imposed fields.
noise floor test, not a deformation testadd imposed rigid translations from a micrometer stage
Stereo-DIC Challenge 1.0stereo camera images of a speckled plate with 3D features on a 2-axis stage; two stereo rigs (16 and 35 mm lenses)18 steps x 5 images per camera, 2 rigs, plus calibration images and simulated imagesstage encoder translation (+-300 nm) of +-20 mm, plate shape from laser scan and CMM; real, plus simulated images with commanded motionboth imagesdirectnot stated on the iDICs pageunknownwaits for a tool
Waits for stereo calibration that carries the 3D truth into each camera: a stage-translated plate (rigid) and simulated 3D tension samples; public Drive folder, no licence stated
Rigid motion only, and stage axes are slightly misaligned with the specimen (paper gives the correction). Needs stereo calibration.Use the simulated images with a known non-rigid morph of the plate surface for deformable truth.
Stereo-DIC Challenge 2.0stereo DIC images (content unverified)unverifieddirectnot statedunknownwaits for a tool
Waits for stereo calibration, as 1.0: synthetic and real tensile stereo images in a public Drive folder, truth a 3D surface field; no licence stated
Current challenge; only the Drive link was seen.Stage-commanded rigid motion and simulated images with a known surface morph, as in Stereo-DIC 1.0.
Stereo4D (dynamic 3D tracks mined from internet stereo video)internet VR180 stereo videos turned into dynamic 3D reconstructionslarge scale (count unverified)pseudo truth fused from pose, stereo depth and tracking methods (not independent)one imagedirectunverifiedunknownwaits for a tool
internet stereo video with method-made 3D tracks: waits for point tracking; the tracks are not independent truth
arXiv Dec 2024, CVPR 2025.use only as training data; evaluate on human-verified tracks
Strecha multi-view (fountain-P11, Herz-Jesu-P8, castle-P19 and others)high-resolution photos (2D) against LIDAR-scanned triangle models (3D) of building facades and a fountain6 scenes, 8-30 images eachcamera matrices calibrated against LIDAR ground truth (Strecha et al. CVPR 2008)both imagesdirectnone stated; cite the paperunknownwaits for a tool
photos against LIDAR meshes of facades: waits for a 2D/3D pose error
real; ground-truth meshes downloadable only for fountain-P11 and Herz-Jesu-P8 on the page
Sugar4D (4D LiDAR sugar beet with temporally consistent leaves)terrestrial LiDAR point clouds of growing plants over 16 time points768 point clouds of 48 plantsmanual temporally consistent organ instance labels (675 tracked leaves)both imagesdirectCC BY 4.0 (bonndata)openwaits for a tool
LiDAR point clouds of growing plants, no images: waits for mesh scoring
Sci Data 2026. Point clouds, not images; growth deformation.
SUN3Dindoor RGB-D video, two views254 sequencescamera poses refined by bundle adjustmentboth imagesdirectnone stated (page, data folders, paper)askin the corpuspriority 3
In the corpus, ranked apart: one pair of frames half a second apart from each of 38 of the 201 places with poses, 20 test and 18 development, points carried by the sensor's depth and the provider's poses (automatic); no licence stated (ask).
Indoor two-view benchmark (OANet split). Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Image matching survey 2025).
Syn4D (multiview synthetic 4D dataset)rendered multi-view videos of dynamic scenes with humanscount unverifiedexact camera motion, depth, dense tracking and SMPL pose from the rendererboth imagesdirectunverifiedunknownwaits for a tool
rendered multi-view video with exact tracks: waits for point tracking
Oxford/NTU/Naver 2026. Synthetic.
Synthetic MS-COCO homography pairs (DeTone 2016 / UDHN protocol)COCO photo crops warped by a random 4-corner perturbationgenerated on the fly, unlimitedexact homography by construction (protocol of DeTone et al., 2016)both imagesdirectCOCO: annotations 'licensed under a Creative Commons Attribution 4.0 License'; 'Use of the images must abide by the Flickr Terms of Use.'askleft out
A protocol warping one photo against itself by a random homography: no second image; COCO photos keep their Flickr terms
A protocol, not a dataset. One image warped: no parallax, light or time change; big domain gap.
Synthetic speckle DIC/DVC generators (StrainNet Speckle 1.0/2.0, Deep DIC, DVC-Simulator)generated 2D speckle pairs and 3D bead volumes deformed by random known fieldsunlimited (StrainNet paper: 36,300 pairs from 363 references x 101 warps)exact imposed displacement; syntheticboth imagesdirectStrainNet: NOASSERTION; Deep DIC (RuYangNU) and FranckLab DVC-Simulator: no licence fileaskleft outpriority 3
generators, not released images: the corpus's own known warps do the same
Generators, not fixed sets: fix seeds and publish the generated test split to make it a benchmark.
SynthVerse (synthetic point tracking dataset)rendered videos (film-style, embodied manipulation, navigation, articulated objects) with dense tracks48K sequences, 5.8M training and 59K test framesexact tracks from the rendererboth imagesdirectunverifiedunknownwaits for a tool
rendered video with exact tracks: waits for point tracking
Synthetic.
T-LESS (BOP)RGB-D and RGB photos (2D) of texture-less industrial parts against their CAD and reconstructed models (3D); 6-DoF pose30 objects, 20 test scenes; Primesense, Kinect 2 and Canon images6D poses from a calibrated turntable/marker-board setup with manual refinement (Hodan et al. 2017)both imagesdirectCC BY 4.0openwaits for a tool
RGB-D photos against CAD models: waits for a 2D/3D pose error
real test images; symmetric look-alike parts make it the hardest classic photo-to-CAD set
Taipei-VLoc UAV visual localization with temporal shifts (Kaggle)continuous low-altitude UAV frames vs older georeferenced satellite mapsunverifiedposition truth type unverifiedaccountUnknownunknownneeds an account
the download needs an account
includes seasonal and structural changes between query and mapknown flight GNSS logs as truth if present; else 20 manual point pairs per flight
Tanks and Templesvideo frames (2D) against industrial laser scans (3D) of outdoor and indoor scenes21 scenes (training 7, intermediate 8, advanced 6) with laser ground truth for training sceneslaser scans (FARO) aligned to image reconstruction; camera poses from COLMAP aligned to the scanboth imagesdirectpage says CC BY 4.0 but its licence grant limits use to non-commercial research (contradictory)non-commercialwaits for a tool
video frames against laser scans: waits for a 2D/3D pose error
real; poses are SfM aligned to scans, not measured independently; licence text contradicts itselffor 2D-3D, use training scenes: perturb COLMAP poses and score against scan-aligned poses
TAP-Vid (DAVIS, Kinetics, RGB-Stacking, Kubric) + RoboTAPpoint tracking through real videos (DAVIS, Kinetics, robot videos) and synthetic scenes30 DAVIS + 1000 Kinetics + 50 RGB-Stacking videos; RoboTAP robot videos; any frame pairhuman-clicked point tracks with flow-assisted interpolation, checked by annotators (DeepMind); synthetic tracks exactboth imagesdirectTAP-Vid annotations, RGB-Stacking and RoboTAP videos: 'Creative Commons BY license'; DAVIS and Kinetics videos under their own licencesopenwaits for a tool
video with human point tracks: waits for point tracking
Kinetics videos must be fetched from YouTube (some gone); sparse points with occlusion flags.
TAPVid-3D3D point tracking in real videos: Aria Digital Twin, Panoptic Studio, Waymo Open4,569 clips, 25-300 frames, ~50 tracks per clipmetric 3D tracks from digital-twin geometry (ADT), multi-view triangulation (Panoptic), lidar + boxes (Waymo) (DeepMind and others)both imagesaccountannotations and code 'released under a slightly modified Apache 2.0 license'; users must agree to the Waymo Open, Aria Digital Twin and Panoptic Studio termsaskneeds an account
the download needs an account
Generated by scripts from source datasets; 2D projections usable as point truth.
TAPVid-MV (multi-view 3D point tracking benchmark)synchronized multi-camera videos (robotics, people, driving, synthetic), 3D point tracks284 sequences, 1,142 camera streams, 109,769 trackstracks from depth, LiDAR, SLAM/SfM, meshes or simulation, each visually verified by annotatorsboth imagesdirectunverifiedunknownwaits for a tool
multi-camera video with 3D point tracks: waits for point tracking
Google DeepMind 2026.
TartanAir (v1 and v2)synthetic Unreal Engine environments, camera trajectories; v2 adds lidar, fisheye, events1M+ frames (v1); v2 larger (multi-TB)exact optical flow, depth and poses by rendering (CMU AirLab)both imagesdirect: anonymous S3 reads from airlab-cloud.andrew.cmu.edu (tartanair_tools download_training.py) or the ungated Hugging Face mirror theairlabcmu/tartanairCC BY 4.0 (theairlab.org dataset page, the provider's page); BSD-3-Clause on the Hugging Face mirroropenin the corpuspriority 3
In the corpus, ranked apart: v1, one pair of consecutive frames for each of its 15 independent environments (three same-world duplicates left out), 10 test and 5 development, scored on the rendered flow outside the occlusion mask.
Synthetic domain gap; Hugging Face mirror says BSD-3-Clause (conflicts with site).
TartanGround (ground robot simulation dataset)photorealistic simulated ground-robot stereo cameras, LiDAR910 trajectories in 70 environmentsground-truth poses, depth, optical flow, disparity from the simulatorboth imagesdirectunverifiedunknownleft outpriority 3
same simulated environments as TartanAir, which ships the flow: build TartanAir instead
CMU AirLab 2025. Synthetic.
TBBR (thermal bridges on building rooftops, RGB-thermal UAV)UAV RGB and thermal infrared images of roofs, same scene3,000 train / 300 test pairs samplednear-aligned by the dual camera rig (unverified)directunknownunknowntruth to make
Found on Zenodo (7022736, CC BY 4.0, 68.5 GB): RGB, thermal and height come as one co-registered stack, so a known warp would make the truth
Download page not confirmed (Zenodo search failed).known homography or smooth warp applied to aligned pairs gives exact truth
TomoBank dynamic and in situ sets (APS, SLS)Synchrotron tomography: dynamic (foam, fuel cell, methane hydrate, interlaced) and in-situ (fatigue-corrosion, high pressure, rock permeability); same sample over timeTime series per sample (unverified counts)Noneone imagedirectunverified (no licence stated on the page)unknowntruth to make
synchrotron projections of samples changing over time, to reconstruct first; no truth, a known warp would make it
Mostly raw projections for reconstruction research; reconstruct first.Reconstruct, then known smooth warp per chunk; foam/fatigue series give real deformation without truth, so pair with a DVC baseline
Triturus newt skull 3D landmarks (Ivanovic & Arntzen)newt skulls of 8 Triturus species, micro-CT surface models, between specimens177 specimens3D landmarks on micro-CT skull models by the authorsboth imagesdirectCC0 1.0openleft out
only landmark coordinates are released; the micro-CT scans are not in the record
Fails: CT scans and meshes are not in the record, only coordinates. Typical of morphometrics data where scans sit on MorphoSource or nowhere.Pick specimens with public MorphoSource scans and reapply the landmark protocol; two people landmark 10 skulls
UEOF (underwater event-camera optical flow)ray-traced underwater RGBD sequences converted to event streamscount unverifieddense ground-truth flow, depth and camera motion from the rendererboth imagesdirectunverifiedunknownleft out
Event streams with flow, not image frames (the ray-traced RGB is not listed as released); CC BY-NC-SA 4.0
Synthetic, 2026.
UVDocsemi-synthetic: captured real paper geometry re-textured with documents; separate benchmark settraining set + benchmark setphysically captured 3D grid and unwarping grid (Verhoeven, Magne, Sorkine-Hornung, ETH)both imagesdirect: https://igl.ethz.ch/projects/uvdoc/UVDoc_benchmark.zipnone on the hosting page; MIT on the UVDoc-Dataset and UVDoc code repositoriesaskin the corpuspriority 2
UVDoc_benchmark.zip (0.78 GB, ranges accepted) holds 50 samples, each a photograph-like image of a real deformed sheet re-textured with a document, its flat texture, the measured 2D and 3D grids and the UV map (zip listing). The deformation is measured from real paper but the image is composited through it, so the truth is exact by construction (synthetic, ranked apart). The ETH page that hosts the files states no licence (site copyright only); the UVDoc-Dataset GitHub repository, whose README says it contains the dataset, is MIT.
Real geometry with composited texture.
VeRLab non-rigid RGB-D benchmark (GeoBit, GeoPatch, DEAL, DALF)deforming real objects (cloth, paper) under light change: Kinect 1, Kinect 2, simulated cloth6 Kinect1 objects, 5 Kinect2 objects (light/medium/heavy), simulated sequences; each frame vs referenceKinect1: ~50 hand-annotated keypoints per image; Kinect2: 80 points from motion capture; simulation exact; plus dense TPS warps refined photometricallyboth imagesdirectCreative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) Licensenon-commercialleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): not biomedical; its licence is not shown to allow commercial use (Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) License). It was proposed as: Deforming cloth, paper and bags with about 51 hand-placed points an image (Kinect 1, 6 objects) and 80 motion-capture points (Kinect 2, 5 objects): real non-rigid truth, 5-12 px left after an affine fit, but only 6 people-made objects; CC BY-NC 4.0
DALF's non-rigid evaluation uses this set. A separate 'Non-rigid World' set was not found (unverified).
ViLD UAV-to-satellite image matching and geolocalizationUAV frames vs satellite imagery of the same groundtwo flights (Flight 09 and 10) with train/val/test splitsground-truth coordinates in Lambert 93 and UTMboth imagesdirectCC BY 4.0openneeds a request
Password-protected archive sent on e-mail request; its truth is frame geolocation, not correspondences; 48.5 GB
paper arXiv 2512.02737
VIPER (Playing for Benchmarks)synthetic driving, riding and walking video from GTA V, many weathers and times250k+ frames with flow for every frame; test truth hiddendense optical flow captured from the game engine (Richter, Hayder, Koltun)both imagesdirectThe data we provide is for research and educational purposes only. Commercial use is prohibited.non-commercialneeds a requestpriority 3
files are listed only after a click-through agreement
Photorealistic game imagery; authors ask users to buy GTA V.
Virtual KITTI 2synthetic Unity clones of 5 KITTI sequences with weather and camera variations, stereo5 sequence clones x ~10 variationsexact forward/backward flow, scene flow, depth and poses by rendering (Naver Labs Europe)both imagesdirectmay be used for non-commercial purposes only and is subject to the Creative Commons Attribution-NonCommercial-ShareAlike 3.0non-commercialleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): not biomedical; its licence is not shown to allow commercial use (may be used for non-commercial purposes only and is subject to the Creative Commons Attribution-NonCommercial-ShareAlike 3.0). It was proposed as: Exact rendered flow on Unity clones of five KITTI sequences; CC BY-NC-SA 3.0
Matched real KITTI counterparts allow sim-to-real tests.
WAAM co-registered in-situ and ex-situ data (ORNL)X-ray CT of a welded block co-registered with in-situ process signals1 buildspatiotemporal registration by the authorsdirectunverifiedunknownleft out
CT against in-situ process signals, not image to image
CT to sensor data, not image to image.CT to process signals, not image to image: out of scope.
Walnut cone-beam photon-counting CT datasetmicro photon-counting CT of walnuts, two energy thresholds, four overlapping bed positions per walnut, raw projections15 walnuts x 4 bed positions x 2 thresholdsnone; known bed geometrydirectCC BY 4.0opentruth to make
raw cone-beam projections to reconstruct first; no truth beyond the known bed geometry, whose overlaps give a stage translation
Energy thresholds share projections (aligned by construction); bed positions overlap, so stitching gives a known-translation check.bed-position overlaps: known stage translation as rigid truth; energy images: known warp on one threshold
WarpDocreal photos of 1,020 printed documents warped six ways (fold, curve, random, rotate, incomplete, perspective)1,020 photo / digital page pairsdigital source page per photo; no correspondences (Xue et al., NTU)one imagedirectnone stated on the pageunknowntruth to make
1,020 photos of printed pages warped six ways against their digital pages, no correspondences, no licence stated: points by people (or DocAligner flow checked by them) would make it
Fails strict criterion; DocAligner derived 800 flow-annotated pairs from it.Dense flow by DocAligner, checked with hand points on a subset.
Waymo Open Dataset (perception)5 cameras (2D) against 5 LiDARs (3D), US drivingabout 2,000 segments of 20 scalibrated camera-LiDAR extrinsics and vehicle posesboth imagesaccountWaymo Dataset License Agreement for Non-Commercial Usenon-commercialneeds an account
the download needs an account
real; Google account and licence acceptance neededHand-check non-box movers on a sample of sweeps.
XYZ-IBD (XYZ Industrial Bin-picking, BOP-Industrial)multi-view RGB-D images (2D) of industrial parts in bins against CAD (3D); bin-picking posereal test scenes plus PBR training (counts not read)millimetre-accurate 6D annotations by XYZ Robotics (method not read)both imagesdirectCC BY-NC-SA 4.0 (custom licence on request)non-commercialwaits for a tool
RGB-D images against CAD: waits for a 2D/3D pose error
real; heavy clutter and occlusion
YFCC100M landmark pairs (OANet split)internet photos of landmarks, two views4 test scenes, ~4,000 pairsSfM pseudo ground-truth poses (COLMAP)both imagesdirectCreative Commons per photoopenleft out
Truth is SfM relative pose only, no depth: it checks an epipolar line, not where a point goes; MegaDepth covers this setting with depth
Standard two-view matching benchmark; truth automatic. Round 4 count: listed in 1 of the 38 round-4 surveys and benchmark papers (Image matching survey 2025).
Zeolite multimodal synchrotron datamulti-resolution micro-CT and XRD-CT of one zeolite fragment1 sample, several resolutionsnoneone imagedirectunverified (Zenodo entries)unknowntruth to make
micro-CT and XRD-CT volumes of one fragment, no truth; stage coordinates between resolutions would give a rigid one
One sample.stage coordinates between resolutions give a rigid truth
ZTF public data releases (single-epoch science images)northern sky, ZTF g/r/i CCD images, same field over hundreds of nights, ground-basedhundreds of epochs per field: effectively unlimited pairsper-image WCS from a SCAMP fit to Gaia reference stars (G 12-18) with 4th-order distortion, by IPAC (Masci et al. 2019); early residuals 45-85 mas per axisboth imagesdirect: https://irsa.ipac.caltech.edu/ibe/data/ztf/products/sci/ (FITS, no account)none stated on IRSA's ZTF pageaskleft out
Left out on 2026-10-27 under Valentin's rule that the corpus takes only relevant, open sets (an open licence with commercial use, a truth made or checked by people, biomedical, a setting the corpus lacks): not biomedical; its licence is not shown to allow commercial use (none stated on IRSA's ZTF page). It was proposed as: IRSA's IBE directories serve the science images without an account (latest release DR24, 2026-01-22); IRSA's ZTF page states no licence or terms. The truth would be IPAC's per-image WCS fitted to Gaia stars by its pipeline (provider, not people). One camera and one field: after the corpus's alignment little beyond the WCS error may be left, so the no-registration line must be measured before building.
Seeing, clouds, satellite trails and non-linear field distortion: a real 2D non-affine star-field problem.Cross-match Gaia DR3 on both frames: hundreds of exact control points per pair; hold out half
Programs that can make a truth

Where a dataset has no ground truth, or a weak one: segmenters, landmark and keypoint finders, annotation tools, simulators. A truth made by a segmenter is a weaker check, and it favours a method that uses the same program.

ProgramWhat it can makeLicenceNotes
3D Slicer MarkupsHuman point lists on any 3D/2D image in linked views, saved as .mrk.json/.fcsv; base for two-reader landmark protocols3D Slicer licence (BSD-style, commercial use allowed)openMost practical tool for 20 landmarks x 10 pairs by two readers.
AALNet33 landmarks on posteroanterior cephalograms (2D X-ray)weights CC BY-NC 4.0 (gated on Hugging Face); code licence not asserted on GitHubnon-commercialSingle-hospital training data.
Abaqus (Dassault Systemes SIMULIA)Commercial finite element solver for biomechanical deformation with known displacementcommercial licence (paid)askOnly if a team already holds a licence; open solvers cover the need.
airlabPyTorch registration library whose transformation models can generate known random warps on 2D/3D imagescode Apache-2.0openMaintenance status unverified.
ANTs displacement field toolsSimulateDisplacementField makes random smooth fields; antsApplyTransforms and antsApplyTransformsToPoints warp images and landmarks exactlycode Apache-2.0openITK (Apache-2.0) offers the same transforms programmatically.
ArtiSynthCombined rigid and finite element biomechanical modelling for known deformations'freely available under the 2-clause BSD license'openJava toolkit.
AutoAFIDs32 anatomical fiducials (AFIDs) on T1w brain MRI, BIDS appcode MIT; models fetched by the app, no separate weights licence written (unverified)openAFIDs is a validated human protocol, so automatic points can be checked by people with the same protocol.
AutoMorphRetinal vessel, artery/vein and optic disc labels on colour fundus photographs; vessel junctions give landmarkscode Apache-2.0; weights licence not writtenunknownFor 2D retina registration where FIRE-style landmarks are missing.
BiomedParse (v1 2D, v2 3D)Text-prompted segmentation and detection across many modalities (CT, MR, ultrasound, X-ray, pathology and more)code Apache-2.0; weights CC BY-NC-SA 4.0 (Hugging Face microsoft/BiomedParse, gated)non-commercialLicence trap: Apache code, non-commercial gated weights.
BrainSegFounderBrain tumour and stroke lesion labels on MR (fine-tuned on BraTS, ATLAS)code GPL-3.0; weights on Google Drive, no separate licence writtenunknownPretrained on UK Biobank (terms unverified); lesion labels are rarely a registration check, low priority.
CADSLabels for 167 structures on CT (whole body)code Apache-2.0; weights in variants: reference = 'CADS customized model license' (plus EXACT09, LOLA11, SLIVER07, EMPIRE10 terms), research = CC BY-NC-SA 4.0, open = CC BY-SA 4.0 (commercial allowed, share-alike), CADS-region = customized licenceopenPick the 'open' variant for commercial use (weaker on some head-neck structures). The reference model was trained on EMPIRE10, a registration benchmark: leakage risk if EMPIRE10 pairs are scored.
Cellpose (incl. Cellpose-SAM)Cell and nucleus masks in 2D/3D microscopy; centroids give landmarks for microscopy pairscode BSD-3-Clause; weights bsd-3-clause on Hugging Face mouseland/cellpose-sam, but README: 'All Cellpose models are trained on data that is licensed under CC-BY-NC'askLicence trap: permissive code and weights tag, non-commercial training data.
CephViT (NLM)Cephalometric landmarks on lateral cephalograms (2D X-ray)README licence 'OpenMDW 1.1' (terms not read, unverified)openCheckpoint is encrypted with a key published in the README; confirm OpenMDW terms before use.
corrField + Foerstner keypoints (Lung250M-4B repo)Automatic dense keypoint correspondences between paired lung CT scans (pseudo truth used for Lung250M-4B), Foerstner interest points, also a grand-challenge algorithmcode Apache-2.0openCircularity trap: corrField-style methods score unfairly well on corrField truth; use only with human verification of a sample.
COSTA / CESARCerebrovascular labels on TOF-MRAcode Apache-2.0; pretrained CESAR weights on Zenodo 10957925 CC BY 4.0openThe COSTA dataset itself is restricted access; the weights are open.
CrossKEYLearned 3D keypoint detector and descriptor matching MR to intraoperative ultrasoundcode MIT; weights licence not writtenopenOne of the few multimodal 3D keypoint tools; check weights availability.
CT-FMCT foundation features (embeddings) for finding corresponding anatomy, plus a fine-tuned SegResNet segmentercode MIT; weights Apache-2.0 (Hugging Face project-lighter/ct_fm_feature_extractor and ct_fm_segresnet)openOpen weights; features can rank candidate correspondences for human checking.
CVATWeb annotation of 2D points and skeletons (and video frames) by teamscode MITopenSelf-hosted avoids the commercial cloud; fits video and 2D pairs.
deeds (deedsBCV)Dense discrete deformable registration for CT/MR; proposes candidate correspondences for people to verifycode MITopenCandidate generator only; never the truth itself for methods of the same family.
DeepDRRPhysics-based X-ray simulation from CT with known C-arm geometry: exact 2D/3D truth with realistic imagescode GPL-3.0openMore realistic than plain DRRs; GPL only on redistribution.
DeepRegRegistration library with random affine and dense-field augmentation that yields known transformscode Apache-2.0openTensorFlow based.
DiffDRRDigitally reconstructed radiographs from CT at known poses: exact 2D/3D registration truthcode MITopenSimple realism; good for 2D/3D tracks.
DINOv2Dense self-supervised features for matching points between 2D images or slices (retina, microscopy, X-ray), basis of DINO-Reg and RoMacode Apache-2.0; weights Apache-2.0 (Hugging Face facebook/dinov2-base)openFully open feature backbone.
DINOv3Dense self-supervised features (successor to DINOv2) for correspondence searchDINOv3 License: royalty-free, commercial use not barred; must show 'Built with DINOv3' when distributed, no military use; Hugging Face access gated (manual)askCommercial use allowed with conditions, but access needs a manual request.
elastix / transformix point mappingMaps landmark sets through a transform (transformix -def) and registers with landmark pairs (CorrespondingPointsEuclideanDistanceMetric)code Apache-2.0openTransfers landmarks to the second image before a person corrects them.
FastSurferWhole-brain parcellation (DKT atlas) and cortical surfaces on T1 MRcode Apache-2.0; weights from the project releases under the repo licence; surface pipeline needs a FreeSurfer licence fileopenOnly the surface module needs the FreeSurfer licence file (free, from a registration form); the asegdkt segmentation does not (per README module list).
FEBioNonlinear finite element biomechanics: realistic tissue deformation (breast compression, organ shift) with known displacementcode MIT (GitHub febiosoftware/FEBio)openNeeds meshes from segmentations; physically plausible warps.
Fiji: BigWarp + Landmark CorrespondencesInteractive landmark pairs on two 2D/3D images with live thin-plate warp (BigDataViewer, very large images); SIFT feature points and landmark-based transformsBigWarp GPL-2.0; mpicbg (Landmark Correspondences, SIFT) GPL-2.0openBest tool for large microscopy volumes; exports landmark CSV and transforms.
ITK MaskFeaturePointSelectionFilterSelects distinctive 3D feature points (high local variance) inside a mask, used with block matchingcode Apache-2.0openBuilding block for automatic landmark candidates; pair with block matching then a human check.
ITK-SNAPSemi-automatic 3D labels (active contours) and manual annotation on CT/MRcode GPL-3.0openFree desktop tool; GPL only binds redistribution.
KeyMorph / BrainMorphLearned corresponding keypoints on brain MR (many contrasts), closed-form transform from themcode MIT; BrainMorph weights on Cornell Box and IXI weights in releases, no weights licence writtenunknownKeypoints are learned by registration, so not independent truth; candidates only.
Label StudioWeb labelling of 2D keypoints, boxes, maskscode Apache-2.0open2D only.
LightGlueFast 2D keypoint matcher; pairs with DISK, ALIKED or SuperPoint keypointscode and LightGlue weights Apache-2.0; DISK Apache-2.0; ALIKED BSD-3-Clause; SuperPoint 'a different, restrictive license' (Magic Leap)openLicence trap: use DISK or ALIKED + LightGlue; SuperPoint + LightGlue is non-commercial.
LoFTRDetector-free dense 2D matching (keypoint pairs) between imagescode Apache-2.0; weights on Google Drive, no separate licence written; indoor model trained on ScanNet, outdoor on MegaDepth; the optimal-transport variant needs Magic Leap SuperGlue code ('strict licence', not shipped)openLicence trap: use the dual-softmax outdoor model; the optimal-transport variant pulls non-commercial SuperGlue code, and ScanNet terms are non-commercial (unverified effect on weights).
lungmaskLeft/right lung and lobe labels on chest CT (R231, LTRCLobes, fused)code Apache-2.0; weights from the project releases, no separate licence writtenopenLTRCLobes was trained on an LTRC subset (data terms unverified). Lobes are a good check for inhale-exhale CT.
MatchAnythingCross-modality 2D matcher (trained for medical, remote sensing, thermal) giving keypoint pairs across modalitiesProject Registration License (PRL) v1.0: free for research and personal use; any company or organisation project, commercial or not, 'is permitted free of charge, provided that the Project Use is registered before such use begins' (Google Form); Hugging Face eloftr card apache-2.0askLicence trap: free but a registration form is required before any organisational use.
MedLAM / MedLSAMFew-shot anatomical point localisation on CT (place a point once on a template, find it in new scans)GitHub licence field Apache-2.0, but README: 'This project is under the CC-BY-NC 4.0 license'; MedLAM checkpoint on Google Drivenon-commercialLicence trap: the repo metadata and README disagree; the README is non-commercial.
MedSAM2D box-prompted segmentation across CT, MR, X-ray, ultrasound, endoscopy, fundus, pathology (slice by slice for 3D)code Apache-2.0; weights Apache-2.0 (Hugging Face wanglab/medsam-vit-base card)openNeeds a person to draw boxes: speeds human labelling, not automatic. Training data mix had varied licences (unverified per dataset).
MedSAM2Prompted 3D and video segmentation (CT, MR, PET, ultrasound, endoscopy video), propagates a mask through slices or framescode Apache-2.0; weights CC BY-SA 4.0 (Hugging Face wanglab/MedSAM2)openCommercial use allowed but share-alike on derived weights; outputs (labels) are not weights.
MeVisLabVisual programming for custom landmarking and review toolsproprietary: free SDK with limits (non-commercial), paid non-commercial licence, paid commercial Professional SDKaskCommercial use needs a paid licence.
MITK WorkbenchPoint set annotation and labels on 3D/4D images, point-based registrationcode BSD-3-ClauseopenAlso hosts nnInteractive.
MONAI Auto3DSegTrains an ensemble segmenter for any 3D modality from your labelled imagescode Apache-2.0; no pretrained weightsopenAlternative to nnU-Net; needs labels to train.
MONAI LabelAI-assisted labelling server for 3D Slicer, OHIF, QuPath (DeepEdit, segmentation apps)code Apache-2.0openCheck the licence of any model an app downloads (some MONAI bundles are non-commercial).
MONAI Model Zoo bundlesPer-bundle segmenters: whole-body CT (104 classes, TotalSegmentator data), spleen, pancreas, renal, brain, prostate MR, ventricles, pathology nuclei; also a valve landmark regressorper bundle: most bundles Apache-2.0 for code and weights; vista3d, vista2d, maisi_ct_generative weights NVIDIA non-commercial; multi_organ_segmentation 'Private Use, Non-Commercial, Non-Reverse-Engineering License'; each bundle's data_license.txt lists third-party training data (renalStructures_CECT data CC BY-NC-SA 4.0)askCheck LICENSE and docs/data_license.txt of each bundle before use; wholeBody_ct_segmentation is Apache-2.0 and trained on TotalSegmentator data.
MONAI Rand3DElastic / RandDeformGridRandom elastic and grid deformations of 2D/3D images with known fieldscode Apache-2.0openSame caveat as TorchIO.
MONAI valve_landmarks bundle10 valve-insertion landmarks on 2D cine cardiac MR (2-, 3-, 4-chamber views)code and weights MIT (bundle LICENSE)openTraining data private; useful for 2D+t cardiac MR frame pairs.
MOOSEWhole-body organ, bone, muscle, vessel labels on CT (PET/CT studies)code Apache-2.0; models 'CC BY 4.0' (README badge); a commercial version is offered through ZentaopenGood for PET/CT and longitudinal whole-body CT pairs.
MRSegmentatorLabels for 40 classes on MR and CT of abdomen/pelvis/thorax, plus 10 MR body-composition classescode Apache-2.0; weights auto-downloaded from the project, no separate weights licence written (repo Apache-2.0)openUseful for MR-CT and MR-MR abdominal pairs; nnU-Net based.
napari + affindernD points layers on large arrays; affinder fits an affine from clicked correspondence points between two imagesnapari BSD-3-Clause; affinder BSD-3-ClauseopenGood for microscopy and OME-Zarr pairs; nnInteractive plugin adds labels (non-commercial weights).
NaviAirwayAirway tree labels on chest CT (branch points give landmarks)code MIT; trained on EXACT09 and LIDC-IDRI; released weights not confirmedunknownTotalSegmentator lung_vessels (open) already gives airways; NaviAirway reaches finer branches (unverified).
NeuroglancerWeb viewer for huge volumes (precomputed, N5, Zarr) with point annotations shared through the URL statecode Apache-2.0openFor large-volume pairs where desktop tools cannot load the data.
NiftySimGPU nonlinear finite element soft-tissue simulation (TLED) for known deformationsBSD-3-Clause on the dockerised copy github.com/eloygarcia/niftysim; original licence unverifiedopenOld CUDA code; licence of the original distribution not seen.
nnInteractive (+ napari plugin, Slicer, MITK)Fast human-in-the-loop 3D labels from points, boxes, scribbles and lasso on any 3D modalitycode Apache-2.0; checkpoint 'Creative Commons Attribution Non Commercial Share Alike 4.0' (README); napari-nninteractive plugin Apache-2.0non-commercialLicence trap: the tool is Apache but the only official checkpoint is CC BY-NC-SA 4.0; the library prints the licence at load.
nnU-Net (framework)Trains a segmenter for any modality from your own labelled images, then labels both images of every paircode Apache-2.0; no pretrained weights in v2openThe engine behind TotalSegmentator, MRSegmentator, CADS, MOOSE; truth only as good as the training labels.
nnU-Net v1 pretrained modelsLabels for the Medical Segmentation Decathlon and other challenge tasks (CT, MR)weights CC BY-NC 4.0 (Zenodo record)non-commercialLicence trap: the code is Apache-2.0 but the released v1 weights are non-commercial.
OHIF ViewerWeb DICOM viewer with measurement tools for remote readerscode MITopenPoint probes exist but 3D landmark workflow is basic; good for distributed reading.
OmniGlueGeneralisable 2D keypoint matching (SuperPoint keypoints + DINOv2 guidance)code Apache-2.0; uses SuperPoint weights from rpautrat/SuperPoint (MIT re-implementation) and DINOv2 (Apache-2.0); OmniGlue weights licence not written separatelyopenUses the MIT SuperPoint re-implementation, not Magic Leap's non-commercial one.
plastimatch synth-vf / warpsynth-vf creates synthetic vector fields; warp applies them to images, structures and 3D Slicer landmark filesBSD-style licence (plastimatch.org)openRadiotherapy-oriented; handles DICOM-RT structures.
QuPathPoint and region annotations on whole-slide images; cell detection gives centroidscode GPL-3.0openFor histology section pairs (restains, serial sections).
RoMaDense 2D matching with DINOv2 features, robust to appearance changeREADME: 'All our code except DINOv2 is MIT license', DINOv2 Apache-2.0; weights licence not written separatelyopenIndoor variant trained on ScanNet (same caveat as LoFTR).
SAM / UAE anatomical embeddings (Alibaba DAMO)Pixel-wise anatomical embeddings that find the corresponding point of a template landmark in another CT (one-shot landmark transfer)code MIT; checkpoints on ModelScope, licence not seen (unverified)unknownStrong candidate generator for CT landmarks; weights licence must be checked on ModelScope.
SAM 2 / SAM 2.1 (Meta)Prompted segmentation and mask tracking in 2D images and video; works on microscopy and endoscopy framescode Apache-2.0; weights Apache-2.0 (Hugging Face facebook/sam2.1-hiera-large)openBase of MedSAM2 and many Slicer/napari tools; fully open.
SAM-Med3DPoint-prompted 3D segmentation on many CT/MR targetscode Apache-2.0; weights on Hugging Face blueyo0/SAM-Med3D with no licence tagunknownTrained on SA-Med3D-140K, a mix of public datasets with their own licences; weights licence not written.
SIFT3D3D SIFT keypoints and descriptors with RANSAC matching on any 3D image (CT, MR)code MITopenOld but open; gives candidate landmark pairs for human checking.
Simul@trophyLongitudinal brain MRIs with known atrophy deformation fields from a biophysical modelBSD-style licence with required citations (Copyright INRIA 2016)openKnown small longitudinal brain changes; good for subtle-change tracks.
Slicer Landmark Registration modulePaired landmark placement on fixed and moving images with live affine/thin-plate warp previewLicense.txt follows the 3D Slicer licence text (BSD-style; first lines seen)openBundled with Slicer; the live warp shows the next landmark's likely position, which speeds pairing (watch for bias).
SlicerMorph (ALPACA, MALPACA)Fast landmarking on 3D models/volumes, semi-landmarks, and automated landmark transfer from one or several templates (ALPACA, MALPACA)code BSD-2-ClauseopenMALPACA transfers a person's template landmarks to new scans; then a person corrects them.
SOFAReal-time finite element simulation of soft tissue for surgical deformation with known fieldscode LGPL-2.1openLGPL allows commercial use when linked.
StarDistNucleus instances in 2D fluorescence and H&E images, 3D models trainablecode BSD-3-Clause; pretrained models: no separate licence; 2D_versatile_fluo trained on a DSB 2018 subset, 2D_versatile_he on MoNuSeg 2018 + TNBCaskMoNuSeg data terms likely non-commercial (unverified); fluorescence model is the safer one.
STU-NetLabels on CT from models pretrained on TotalSegmentator (S/B/L/H sizes), fine-tunable on new taskscode Apache-2.0; weights on Google Drive/Baidu, no separate licence written (repo Apache-2.0)openWeights licence inferred from the repo; mostly useful as a fine-tuning start.
SuperPoint / SuperGlue (Magic Leap)2D keypoints and matches'ACADEMIC OR NON-PROFIT ORGANIZATION NONCOMMERCIAL RESEARCH USE ONLY' (as stated on Hugging Face ETH-CVG/lightglue_superpoint, implied by SuperPoint)non-commercialLicence trap: Hugging Face transformers ports (magic-leap-community/superpoint, lightglue_superpoint) inherit it.
SuPreM / AbdomenAtlas models (JHU)Abdominal organ and tumour labels on CT (25+ classes, pretrained backbones)GitHub LICENSE: 'openly licensed via CC BY-NC-ND' (both SuPreM and AbdomenAtlas repos); Hugging Face MrGiovanni/SuPreM card says apache-2.0askLicence trap: the two pages conflict; treat as non-commercial no-derivatives until JHU confirms.
SynthMorph label-to-image generator (voxelmorph + neurite)Images of arbitrary contrast synthesised from label maps with known random warps; exact multimodal truth by constructioncode Apache-2.0 (voxelmorph, neurite); SynthMorph model files on the FreeSurfer server, licence not written; FreeSurfer build under the FreeSurfer licenceopenGood for multimodal brain tracks; images look synthetic.
SynthSegBrain labels (whole-brain parcellation) on MR of any contrast and resolution, also CT; its generator makes images of any contrast from one label map, so aligned multimodal pairs by constructioncode Apache-2.0; weights via UCL SharePoint, no separate weights licence (repo Apache-2.0); also shipped in FreeSurfer (mri_synthseg) under the FreeSurfer licenceopenContrast-agnostic, so the same label protocol on both images of a multimodal brain pair.
SynthStripBrain masks on any MR contrast, CT, PETFreeSurfer Software License: royalty-free licence to use, modify and distribute, 'including ... incorporate the Software into proprietary programs'; 'designed for research purposes only' (not FDA reviewed)openFreeSurfer licence is permissive (commercial allowed). Full FreeSurfer install needs a free licence key from a registration form; standalone SynthStrip container key need unverified.
TopCoW models (Circle of Willis)13-class Circle of Willis labels on CTA and MRA; centerline skeletons and bifurcations from the CoW masksbest TopCoW dockers (Zenodo 15665435): no licence stated; TopCoW 2024 MRA winner weights (Zenodo 14191592) CC BY-NC 4.0; CoW skeletonization nnU-Net weights (Zenodo 17357136) CC BY 4.0 with code at github.com/fmusio/CoW_Centerline_Extraction (no licence file found)askLicence trap: the winning MRA model is non-commercial; only the skeletonization weights are clearly open. CoW bifurcations make natural landmarks for brain CTA/MRA pairs.
TorchIO RandomElasticDeformationRandom smooth B-spline warps applied to a real 3D image and its labels: exact known displacementcode Apache-2.0openSynthetic truth only tests what the warp model can produce; mix with real pairs.
TotalSegmentator (licensed tasks)Labels on CT/MR: tissue_types(_mr), tissue_4_types, heartchambers_highres, appendicular_bones(_mr), brain_structures, face(_mr), thigh_shoulder_muscles(_mr), coronary_arteries, aortic_sinuses, brain_aneurysmcode Apache-2.0; weights 'Available with a license (free licenses available for non-commercial usage ... For a commercial license contact jakob.wasserthal@usb.ch)'; brain_aneurysm 'CC BY-NC 4.0 license, no commercial license available'non-commercialLicence trap: free licence is non-commercial and obtained through a web form; body stats and aorta report also need it. brain_aneurysm can never be used commercially.
TotalSegmentator (open tasks, CT)Labels on CT: 117 structures (task total) plus lung arteries/veins/airways, body, vertebrae_pp, vertebrae_body + discs, liver segments, head-neck, craniofacial, teeth and more; vertebra centroids from masks give landmarkscode Apache-2.0; weights of the open tasks 'Openly available for any usage (Apache-2.0 license)'openBest open default for CT label truth. vertebrae_body is now in the open list (no licence needed). Tasks marked * on the README were trained on small datasets: less robust. Labels are a weaker check than landmarks.
TotalSegmentator MRI (total_mr)Labels on MR of any sequence: 50 structures (total_mr), plus body_mr, vertebrae_mr, liver_segments_mrcode Apache-2.0; weights Apache-2.0 (listed under 'Openly available for any usage')openSame package as the CT tasks; lets one label set serve MR-MR and MR-CT pairs.
Vertebra localisation (Payer, VerSe)Vertebra centroid landmarks and labels on spine CTcode GPL-3.0; released weights not confirmedunknownTotalSegmentator vertebra masks (open) give centroids too.
vesselFMZero-shot 3D vessel labels on any modality (CTA, MRA, microscopy, etc.)code GPL-3.0; weights 'Open RAIL++-M License and are restricted to research and non-commercial use only'non-commercialLicence trap: RAIL weights with a non-commercial restriction.
VISTA3D / NV-Segment-CTMR (NVIDIA)Automatic and point-prompted 3D labels, 127 CT classes (MR coverage implied by the model name, unverified)code Apache-2.0; weights 'NVIDIA License ... only may be used or intended for use non-commercially ... research or evaluation purposes only' (README: NVIDIA OneWay Non-Commercial License)non-commercialLicence trap: same NVIDIA non-commercial weights clause in the MONAI vista3d bundle and MAISI.
vmtk (+ Slicer VMTK extension)Vessel centerlines, branch splitting and bifurcation points from vessel segmentations (any 3D angiography)BSD licence (vmtk), ITK Apache-2.0, VTK BSDopenSegment vessels on both images (TotalSegmentator, vesselFM, COSTA), extract bifurcations, match, then have a person verify; skan (BSD-3) does the same on 2D/3D skeletons.
vtkOpenSURF3D (3D SURF)3D SURF keypoints and descriptors for CT/MR volumes, used for groupwise landmark matchingcode GPL-2.0openGPL only matters if redistributed; candidate landmarks only.
webKnossosCollaborative web annotation (skeleton nodes, points, labels) of very large EM and microscopy volumescode AGPL-3.0 (github.com/scalableminds/webknossos)openAGPL obligations only if a modified version is served to others.
XCAT phantoms (Duke CVIT)4D computational phantoms with respiratory and cardiac motion: known motion between phases, simulated CT/MR/PET/SPECTlicence by request from Duke; academic fee shown for the XCAT brain phantom only; commercial terms not foundaskExact motion truth, simulated images.