mirror of https://github.com/open-mmlab/mmpose
4.0 KiB
4.0 KiB
3D human pose estimation in video with temporal convolutions and semi-supervised training
Based on the success of 2d human pose estimation, it directly "lifts" a sequence of 2d keypoints to 3d keypoints.
Results and Models
Human3.6m Dataset
| Arch | MPJPE | P-MPJPE | N-MPJPE | ckpt | log | Details and Download |
|---|---|---|---|---|---|---|
| VideoPose3D-supervised-27frm | 40.1 | 30.1 | / | ckpt | log | videpose3d_h36m.md |
| VideoPose3D-supervised-81frm | 39.1 | 29.3 | / | ckpt | log | videpose3d_h36m.md |
| VideoPose3D-supervised-243frm | 37.6 | 28.3 | / | ckpt | log | videpose3d_h36m.md |
| VideoPose3D-supervised-CPN-1frm | 53.0 | 41.3 | / | ckpt | log | videpose3d_h36m.md |
| VideoPose3D-supervised-CPN-243frm | 47.9 | 38.0 | / | ckpt | log | videpose3d_h36m.md |
| VideoPose3D-semi-supervised-27frm | 57.2 | 42.4 | 54.2 | ckpt | log | videpose3d_h36m.md |
| VideoPose3D-semi-supervised-CPN-27frm | 67.3 | 50.4 | 63.6 | ckpt | log | videpose3d_h36m.md |