code for Training Data Generation

[R132]

Question

I am very interested in the Training Data Generation: Part-level pseudo-label module described in your paper.

Could you kindly share the plan whether you will open-source the corresponding implementation code for this part?

Thanks a lot for your great work!

[WuJH2001]

Hi @R132, thank you for your interest! It should be fairly straightforward if you follow this repo: https://github.com/Pixel-Talk/EHM-Tracker

The main things you would need to do are:

1. Adapt the video-based tracking pipeline to work with individual images.
2. Use a stronger SMPL-based model instead of SMPL-X, or alternatively, use the MHR human dataset released by SAM3D-Body and convert the body poses to SMPL-X. The face and hand parameters can still be obtained using the EHM-Tracker repository.
3. Handle the gap between SMPL/MHR parameters and the SMPL-X body pose. You could implement an automated optimization procedure to refine the converted body pose and minimize this gap.

Overall, the whole pipeline should not take too much time to implement.