search

athn-nik / teach

Official PyTorch implementation of the paper "TEACH: Temporal Action Compositions for 3D Humans"
https://teach.is.tue.mpg.de
Other
383 stars 40 forks source link

Body model #26

Closed yonishafir closed 1 year ago

yonishafir commented 1 year ago

Hi, I've noticed your'e using "mmm" body model instead of smplh.

https://github.com/athn-nik/teach/blob/2f81b1b7e00fe279e2b4c3b557799589e8247a6a/teach/transforms/rots2joints/smplh.py#L38

I just want to make sure I understand correctly and you're indeed using it and its kinematic chain.

I'll also be happy for a short explanation of why you choose this instead of smplh.

Thanks.

athn-nik commented 1 year ago

Hey @yonishafir ,

I am using mmm joints to follow the previous works and make sensible comparisons. As the previous papers of Ghosh et. al and L2P, TEMOS use those and are trained on those the comparisons would not be fair.

Notice that those are used for evaluation. The whole training pipeline happens with SMPLH joints-based features. After training, I am using the mmm for valid comparisons and as TEACH can naturally also support SMPL rotations I am using them for visualizations (the model is actually learning body movements and not mmm skeletons as prior work). The conversion, which is made to enable fair evaluation, is being done as in TEMOS using this mapping.