fidopt Interactive view
Registration leaderboards

Photographs of a planar scene (2D), ranked apart

1 datasets, 10 test pairs. A method is ranked here when it has a result on every pair of every one of them.

Read before comparing

  • These correspondences come from a transform the dataset's authors estimated, partly by hand and partly by an automatic fit, and nobody checked them afterwards. These leaderboards are ranked apart until people do.
  • Ranked apart: outside biomedical imaging, kept as a generic 2D registration problem.

Ranking OfficialResearch

#MethodMean rankHPatches
1vismatch matchers · ALIKED + LightGlueGPU2.001.0–6.01
1vismatch matchers · RoMaGPU2.001.0–6.01
1vismatch matchers · SIFT + LightGlueGPU2.001.0–6.01
4vismatch matchers · DISK + LightGlueGPU4.001.0–6.04
5vismatch matchers · XFeatGPU5.001.0–6.05
6vismatch matchers · LoFTRGPU6.001.0–6.06
1 place in that table, 1 is best2/10 cases done, not ranked yetn/a does not apply not run some cases failed

Under each dataset, the method's place in that dataset's table: 1 is best, the darker the better. The mean rank averages its places over the datasets.

The scores, dataset by dataset

Each score is followed by its change against no registration: green when the method did better, red when it did worse.