search

Parskatt / DKM

[CVPR 2023] DKM: Dense Kernelized Feature Matching for Geometry Estimation
https://parskatt.github.io/DKM/
Other
385 stars 28 forks source link

Did you test your DKM in the Aachen Benchmark ? #14

Closed noone-code closed 1 year ago

noone-code commented 1 year ago

Like SuperGlue and LoFTR did

Parskatt commented 1 year ago

Hi, we have currently not tested our method on the Aachen benchmark. There is no specific reason for this rather than focusing on the two-view matching problem rather than localization.

The benchmarks we do test on are closely related to the IMC challenge benchmarks, as those benchmarks have previously been derived from the Phototourism dataset. If I were to add a benchmark it would probably be IMC21 as it was recently made open-source.

Parskatt commented 1 year ago

Some additional context for localization/ vs matching:

image

Basically matching methods do not really improve results on Aachen in general, however I don't think this is due to lack of improvement in the matching methods themselves but rather on how they are evaluated (judging from the results of recent IMC challenges it is clear that LoFTR and follow-ups perform superior to SuperGlue). I therefore don't find it a relevant benchmark.

noone-code commented 1 year ago

Yep, I agree with you for this point.

Basically matching methods do not really improve results on Aachen in general