search

ShenhanQian / GaussianAvatars

[CVPR 2024 Highlight] The official repo for "GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians"
https://shenhanqian.github.io/gaussian-avatars
497 stars 59 forks source link

Weird xyz loss in custom data #53

Open aixiaodewugege opened 1 month ago

aixiaodewugege commented 1 month ago

image The XYZ loss is always 0, and the scale loss is five times larger than in your data. Do you know what could be causing this? I use the monocular camera data for training.

ShenhanQian commented 1 month ago

The xyz_loss stays zero when the offset of Gaussians is very small. You may try enlarging position_lr_init, position_lr_final.

They should be adjusted according to the mesh size in the Euclidian space. Same for scaling_lr. You can have a look at this issue to estimate values for these hyper-parameters.

By the way, what does the result look like in the viewer? Is it at least 3D-consistent?

aixiaodewugege commented 1 month ago

Thanks for your reply! image

Here is my result. I use the R T from monocular camera flame tracker, and set it as camera pose. And I set the flame R and T to ones and zeros. Am I handle it right?