fidoptUVDoc, measured paper geometry with rendered document texture
A flat document texture and its rendering on a measured, deformed sheet of paper, with the source UV map as known correspondence (UVDoc).
Read before comparing (6)
- Truth. Up to 256 interior correspondences sampled from the released UV map, not a network's prediction, only where the UV values are finite and inside the document support. The paper's geometry was captured physically and the document's appearance composited through it: semi-synthetic truth, kept apart.
- Alignment. The pair starts as delivered, with no starting transform: bringing the flat texture onto the sheet is the task, and an affine fitted on the known points would hand every method most of the truth. The same points score the whole registration, global pose included.
- Baseline. As delivered the 10 test pairs score a TRE of 370 px (41 to 1948), rTRE 0.13; 00006, a 2666 x 4166 px texture against a 1008 x 1148 photograph, starts farthest.
- Images. RGB values and the released document mask as delivered, each image on its own grid. Coordinates are image pixels, not paper millimetres.
- Independence. Texture filenames are disjoint, but different pages can come from one publication, and documents rendered on measured shapes are not independent natural-image registrations.
- Licence. Source data are prepared locally. No permission to redistribute the archive or its third-party textures follows from the code repository's licence.
| Accuracy | Smoothness | Cost | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| # | Method | TRE px | Change | rTRE | SDlogJ | Folds % | Time s | RAM GB | GPU GB | Images |
| 11–6 | vismatch matchers · LoFTRGPU | 11.73 ±6.01 | −357.78 | 0.0057 ±0.0031 | 0.020 | 0.00 | 4 | 2.3 | 14.7 | |
| 21–6 | vismatch matchers · ALIKED + LightGlueGPU | 9.94 ±3.98 | −359.58 | 0.0048 ±0.0018 | 0.021 | 0.00 | 4 | 2.6 | 1.7 | |
| 21–6 | vismatch matchers · SIFT + LightGlueGPU | 13.73 ±7.22 | −355.79 | 0.0065 ±0.0034 | 0.020 | 0.00 | 4 | 2.4 | 0.5 | |
| 41–6 | vismatch matchers · DISK + LightGlueGPU | 14.28 ±7.35 | −355.24 | 0.0069 ±0.0037 | 0.021 | 0.00 | 4 | 2.5 | 2.9 | |
| 41–6 | vismatch matchers · RoMaGPU | 12.40 ±6.70 | −357.12 | 0.0062 ±0.0037 | 0.021 | 0.00 | 8 | 3.4 | 6.9 | |
| 61–6 | vismatch matchers · XFeatGPU | 23.32 ±28.53 | −346.19 | 0.0081 ±0.0058 | 0.027 | 0.00 | 3 | 2.5 | 0.5 | |
| 7 | greedyCPUapp | 369.59 ±403.22 | +0.08 | 0.1283 ±0.0732 | 0.068 | 0.00 | 13 | 1.0 | ||
| 7 | NiftyRegCPUapp | 368.98 ±403.24 | −0.54 | 0.1280 ±0.0735 | 0.185 | 0.14 | 12 | 0.4 | ||
| No registration | 369.52 ±403.01 | 0.1283 ±0.0732 | 0.000 | 0.00 | ||||||
The order averages each method's ranks on TRE px, SDlogJ, so the first one is not always the best in every column. ± is the 95% interval of the mean over the 10 pairs. Two neighbours are tied when a paired Wilcoxon test on the shared pairs does not find the upper one better on TRE px (5%): the places a method could hold are then shown under its own. The best value of each column is in colour. Click a row to see its images, a column title to sort by it. Time and memory measured on Laptop.
TRE: distance between matching landmarks. rTRE: the TRE over the length of the image's diagonal, to compare datasets of other sizes and units. SDlogJ: how irregular the deformation is, 0 is perfectly smooth. Folds: share of the image where the deformation folds over itself.
Pair by pair
Click a score to look at that result: the fixed image, and over it the moving image after registration, with the annotations. Underlined scores have their image; for the others you see the pair without the result. A green cell gained against no registration on that pair, a red one lost.
About the data
- Test pairs
- 10 test cases drawn with seed 2026 among source geometries whose identifiers begin 09_, 10 development cases among those beginning 11_, all from the released 50-example benchmark, not its training archive. Geometry identifiers and texture filenames are disjoint across partitions; the prefixes are groupings, not verified separate paper specimens.
- Truth
- landmarks a known deformation Up to 256 interior correspondences sampled from the released dense UV map. The image renderer applies this map to a flat document texture using grid_sample with align_corners=False; flat-image pixel coordinates are UV×texture_size−0.5. Paper geometry was physically captured, while the document appearance is composited through that geometry.
- Licence
- The hosting page supplies no explicit archive data licence. The UVDoc-Dataset code repository is MIT-licensed; linked archive images include third-party textures whose terms are not consolidated in the release. Redistribution/commercial permission for this derivative is not established.
10 pairs for tuning, never tested
00048, 00041, 00031, 00038, 00035, 00049, 00045, 00036, 00032, 00042