search

hongsukchoi / TCMR_RELEASE

Official Pytorch implementation of "Beyond Static Features for Temporally Consistent 3D Human Pose and Shape from a Video", CVPR 2021
MIT License
274 stars 39 forks source link

The output is centered at (0, 0, 0) #5

Open liuhaorandezhanghao opened 2 years ago

liuhaorandezhanghao commented 2 years ago

The input character moves to the left, and the model outputs the action to lock the hips, but does not move to the left. How to solve it?

hongsukchoi commented 2 years ago

Hi @liuhaorandezhanghao,

As many works, TCMR produces a mesh centered at the root joint location per frame and weak perspective camera projection parameters. If you want to translate the mesh, you should obtain the absolute 3D coordinate of the root joint.

  1. It is a little bit different research area, but you can checkout this repo to explicitly estimate the absolute 3D root coordinates. https://github.com/mks0601/3DMPPE_ROOTNET_RELEASE

  2. Alternatively, you can use a fixed focal length and estimated camera parameters of TCMR to get the translation.