Train pose estimation model and trajectory estimation model
Predict 3d human pose with root trajectory from the available 2D pose data
The code for Human3.6M data preparation is the same as PoseAug/DATASETS.md.
Once prepared, the folder will be like:
${PoseAug}
├── data
├── data_3d_h36m.npz
├── data_2d_h36m_gt.npz
├── data_2d_h36m_hr.npz
This is enough for SSL learning (2D pose) and evaluation (3D pose) in H36M dataset.
The code includes three parts:
- Estimation models are adapted from VideoPose3D.
- Augmented projection is developed from PoseAug.
- IK part is developed from video2bvh.
If you find this part useful for your research, please consider citing them.