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
622 stars 51 forks source link

What is the meaning of “refine_init” in displacement estimation? #83

Closed Devoe-97 closed 3 weeks ago

Devoe-97 commented 3 weeks ago

Hello, thanks for this great repo. I have a question about what “refine_init” means in the code. https://github.com/Parskatt/RoMa/blob/370117431ffc5dc000fb46f6e581b74bdb2c3ff8/romatch/models/matcher.py#L397

Parskatt commented 3 weeks ago

This is inherited from DKM. My thought then was that that if the output of the network is std=1 this normalization means that outputs are scaled to about 1/4 of the current pixel size on average. I haven't looked if it has any impact. In general its probably good to init the output of the refiners to be near 0 since theyre applied recursively.

Devoe-97 commented 3 weeks ago

This is inherited from DKM. My thought then was that that if the output of the network is std=1 this normalization means that outputs are scaled to about 1/4 of the current pixel size on average. I haven't looked if it has any impact. In general its probably good to init the output of the refiners to be near 0 since theyre applied recursively.

Thanks!