1、download datasets
data_3d_h36m.npz
data_2d_h36m_cpn_ft_h36m_dbb.npz
you can find it in https://github.com/Vegetebird/MHFormer
2、training
arc = 3,3,3,train video sequence nums=27,eg:
python run.py -e 80 -k cpn_ft_h36m_dbb -arc 3,3,3
python run.py -e 80 -k cpn_ft_h36m_dbb -arc 3,3,3,3
python run.py -e 80 -k cpn_ft_h36m_dbb -arc 3,3,3,3,3
3、camera or video
download checkpoints:https://drive.google.com/drive/folders/1TinLUEQ8C0hbyy0T7eMMgc-V0etz-m1s?usp=sharing
python main.py
4、demo for camera:
https://www.bilibili.com/video/BV1a44y1G7Jr?spm_id_from=333.999.0.0
https://www.bilibili.com/video/BV1JA411G7jn?spm_id_from=333.999.0.0
The project is developed based on VideoPose3d(https://github.com/facebookresearch/VideoPose3D).
Combines relatively key points and trajectories to get global 3D key points，Thanks to Viedopose3d researchers.