search

Parskatt / RoMa

[CVPR 2024] RoMa: Robust Dense Feature Matching; RoMa is the robust dense feature matcher capable of estimating pixel-dense warps and reliable certainties for almost any image pair.
https://parskatt.github.io/RoMa/
MIT License
434 stars 33 forks source link

About training settings #7

Closed sugar-fly closed 10 months ago

sugar-fly commented 10 months ago

Hi Johan,

I really appreciate your great work RoMa. I found that the training strategy of RoMa (i.e., scheduler) is different from DKM. Does using different training strategies help with performance?

Thank you so much for your help!

Parskatt commented 10 months ago

I think it does give a slight improvement, stepping later gives a little boost. The baseline in the paper is already with that improvement compared to DKM. I think its about 0.5 points or so.

sugar-fly commented 10 months ago

Thank you for your answer!