search

JYChen18 / RPMG

[CVPR 2022] Projective Manifold Gradient Layer For Deep Rotation Regression
67 stars 3 forks source link

Apply in multi-loss setting #2

Closed Wuziyi616 closed 2 years ago

Wuziyi616 commented 2 years ago

Hi, thanks for open source this great work! I'm wondering how can apply this loss in a multi-loss training setting. Say in pose regression, we have a rotation loss and a translation loss. Can I just add the translation loss (e.g. MSE) and the RPMG loss together? So the only thing I will have to change is replacing the GT rotation with the GT computed from this line. Is that correct? Have you tried RPMG in a multi-loss setup and are there any special things to care about e.g. the loss scaling? Thanks in advance!

JYChen18 commented 2 years ago

Thanks for your attention to our work. You can find the experiment of camera relocalization in this folder, which is very similar to your setting. The comment here may also help you. As for the loss scaling, you can see the discussion in Section E.3 of our appendix. By the way, the rotation computed from this line is a prediction of the network instead of a GT rotation.

Wuziyi616 commented 2 years ago

Thanks for your prompt reply. That helps a lot!