search

ShenhanQian / GaussianAvatars

[CVPR 2024 Highlight] The official repo for "GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians"
https://shenhanqian.github.io/gaussian-avatars
Other
542 stars 85 forks source link

xyz_loss increases during training #37

Closed 3Der1999 closed 5 months ago

3Der1999 commented 5 months ago

Hello, I'm running training code and find xyz_loss increases during training. Is this as expected? image Could you please explain losses['xyz'] = F.relu((gaussians._xyz*gaussians.face_scaling[gaussians.binding])[visibility_filter] - opt.threshold_xyz).norm(dim=1).mean() * opt.lambda_xyz while threshold_xyz = 1.? I don't really understand it. Thanks!

ShenhanQian commented 5 months ago

Hi, the increase of xyz_loss is expected as the Gaussians shift from the triangle to better capture the geometry of the subject.

You can find an explanation of this term by Position loss with threshold in Section 3.4 of our paper.

Vincento-Wang commented 2 months ago

Hello, I'm running training code and find xyz_loss increases during training. Is this as expected? image Could you please explain losses['xyz'] = F.relu((gaussians._xyz*gaussians.face_scaling[gaussians.binding])[visibility_filter] - opt.threshold_xyz).norm(dim=1).mean() * opt.lambda_xyz while threshold_xyz = 1.? I don't really understand it. Thanks!

code share with me ?