search

aipixel / GaussianAvatar

[CVPR 2024] The official repo for "GaussianAvatar: Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians"
https://huliangxiao.github.io/GaussianAvatar
MIT License
343 stars 22 forks source link

reproduce numbers in paper #14

Open hdjang opened 4 months ago

hdjang commented 4 months ago

Hi, Thanks for sharing great work!

I found that model trained on 'm4c_processed' data using the command in READMD.md does not match the numbers in the paper. It also looks similar for the 'dynvideo_male' case too. Can you share us that what config options were used for training? (such as 'batch_size', 'lpips_start_iter', 'pose_op_start_iter', etc)

Best regards

huliangxiao commented 4 months ago

Hi, we also optimize the test sequence to get refined poses for testing before training the model.

hdjang commented 4 months ago

Ah ha! Thanks for the reply

Could you provide more details on optimizing training and test sequence? Such as loss function, optimization process, and etc.

Or did you mean applying training code to test sequence with SMPL optimization, and just use optimized SMPL pose among the checkpoint to update test sequence?

david-svitov commented 4 months ago

And I would be very grateful if you would publish checkpoints for the other three people from SnapshotPeople in the table