fidopt Interactive view
Photographs of scenes (2D), two viewpoints or times

WxBS

Two colour photographs of one outdoor scene from two viewpoints, most often at two times, by day and night or in two seasons: images 01 (fixed) and 02 (moving) of a WxBS pair (version 1.1).

10 test pairsscenesame patientfixed PHOTO, moving PHOTO2D, distances in pxdelivered alignedresearch only
Look at the images
Read before comparing (9)
  • Scenes. WxBS pairs photographs where more than the viewpoint changes, and names its categories after what changes: G the viewpoint, L the lighting, A the appearance (seasons, snow, a flood), S the sensor. Buildings, squares, a park and landscapes, collected from Flickr and other sources. The setting reads "same patient" for want of a word for one scene seen twice.
  • Pool. Version 1.1's 29 visible pairs show 25 scenes; one pair a scene is kept, the first in the order of sha256("2026:WxBS:<case>"), since two pairs of one scene are not independent. The five WGSBS pairs set a thermal image against a photograph, too few for ten test pairs.
  • Images. As delivered, PNG or JPEG, red, green and blue a pixel, 400 x 600 to 1707 x 1280 pixels; the two of a pair differ in size, framing and camera. Distances are in pixels, prepare.py setting every spacing to 1, and the rTRE divides by the fixed image's diagonal.
  • Truth. The authors clicked 12 to 64 correspondences a pair (24 to 40 on the test pairs) by hand on both images with pixelstitch, evenly over the part of the scene both show, mostly on corners and edges at every depth.
  • Precision. Each point's leave-one-out distance to its epipolar lines has a median of 1.3 px over the 983 visible points, 0.7 to 2.2 px a test pair. It misses an error along the line, to another depth: the points are good to one or two pixels.
  • Alignment. The images come unaligned, each corner at the origin, and get no starting transform: two viewpoints differ by a scale, foreshortening and depth, which no translation or rigid fit from the intensities takes out. Neither image is resampled.
  • Parallax. Points at different depths move by different amounts, and the true correspondence breaks at the edge of every object in front of another: a smooth field only approximates it, and is scored only at the points. Lighting and seasons change what the pixels show, so a method that compares intensities sees little that matches.
  • Baseline. With no registration the 10 test pairs score a TRE of 178 px (90 to 344), rTRE 0.20, more than 50 times the points' own error. WGLBS_charlottenburg and WGLBS_maidan start closest, WGLBS_church farthest.
  • Outside numbers. None: the dataset's benchmark ranks matchers by the share of the correspondences their fundamental matrix explains, and point matchers by their distance, each on the pairs as delivered.
AccuracySmoothnessCost
#MethodTRE pxChangerTRESDlogJFolds %Time sRAM GBGPU GBImages
1vismatch matchers · DISK + LightGlueGPU1 failed74.89 ±48.32−103.590.0841 ±0.05160.8709.2032.12.8
2greedyCPUapp178.39 ±58.62−0.090.1988 ±0.06070.0970.0010.3
2NiftyRegCPUapp176.86 ±58.70−1.620.1967 ±0.06020.4910.4630.2
2vismatch matchers · LoFTRGPU83.05 ±51.86−95.430.0976 ±0.06201.1512.1532.12.2
2vismatch matchers · RoMaGPU30.38 ±12.97−148.090.0346 ±0.01461.8964.0183.26.9
66–8vismatch matchers · SIFT + LightGlueGPU5 failed173.92 ±66.00−4.560.2026 ±0.08331.39017.9132.00.5
76–8vismatch matchers · XFeatGPU211.71 ±234.16+33.230.2294 ±0.24202.0784.1532.20.4
86–8vismatch matchers · ALIKED + LightGlueGPU288.83 ±326.76+110.360.3094 ±0.33832.70513.8932.41.7
No registration178.48 ±58.670.1988 ±0.06080.0000.00

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 pairs drawn with seed 2026 among the 15 scenes outside development, the categories in proportion, one pair a scene. The development cases are the provider's two validation pairs, kyiv_dolltheater2 and petrzin, brought to ten from the scenes left. Five scenes are in neither, and the thermal WGSBS pairs are left out.
Truth
landmarks by experts the dataset's correspondences, 12 to 64 a pair (24 to 40 on the test pairs): points clicked by the authors on both images with pixelstitch, their labelling tool, evenly over the part of the scene both images show, added to and cleaned for version 1.1 (Mishkin et al. 2015; the archive's README)
Licence
none stated
10 pairs for tuning, never tested

WGABS_petrzin, WGALBS_kyiv_dolltheater2, WGALBS_flood, WLABS_dh, WGALBS_rovenki, WGLBS_alupka, WGABS_kremlin, WGBS_kn-church, WGLBS_warsaw, WGALBS_stadium