wenj / GoMAvatar

Offical codes for "GoMAvatar: Efficient Animatable Human Modeling from Monocular Video Using Gaussians-on-Mesh"
https://wenj.github.io/GoMAvatar/
98 stars 10 forks source link

Evaluation on PeopleSnapshot #7

Open JuewenPeng opened 3 weeks ago

JuewenPeng commented 3 weeks ago

If I have downloaded the pretrained models, how to evaluate on PeopleSnapshot? I have tried running this code: python eval.py --cfg exps/snapshot_"$SCENE".yaml --type view. However, the metrics are much lower than reported in the paper, and the size of the rendered human is a little larger than GT.

wenj commented 2 weeks ago

We follow previous work and perform the test-time pose optimization to compensate for the errors due to pose mismatch. To optimize the test poses, run python train_pose.py --cfg exps/snapshot_"$SCENE".yaml. This section provides more details.

JuewenPeng commented 2 weeks ago

Thank you very much. I encountered another problem. It seems that you have performed 'undistortion' operation to the image when preprocessing the dataset and also when loading the dataset during training.

wenj commented 1 week ago

If the camera has a distortion parameter provided by the data, the undistortion will be applied. Otherwise, it will ignore it.