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
598 stars 50 forks source link

Value of `coarse_res` and `upsample_res` for ScanNet #29

Open justachetan opened 7 months ago

justachetan commented 7 months ago

Hi @Parskatt ,

I am trying to run an evaluation of roma_indoor on ScanNet, like what you have used for roma_outdoor in RoMa/experiments/eval_roma_outdoor.py. Could you please tell me what values of coarse_res and upsample_res to use with ScanNet when initializing the model?

Thanks!

Regards, Aditya

Parskatt commented 7 months ago

Hi, I think I used the default values as in roma_indoor. Although it could be the case that I didn't use upsample_preds, I don't quite remember.

justachetan commented 6 months ago

Thanks! It seems the code does use upsample_preds: https://github.com/Parskatt/RoMa/blob/d5adf6562bd87dd6d8313b5a20f030a3c80d5caa/roma/models/model_zoo/__init__.py#L34

Would it be possible to get the parameters used to generate the results? I am getting quite poor results for keypoint matching on the ScanNet test case

Parskatt commented 6 months ago

If you run the model (without keypoints), are you able to get results comparable to our reported results?