The official PyTorch implementation of the 5th IEEE/CVF CVPR Precognition Workshop paper Best Practices for 2-Body Pose Forecasting.
Visit our webpage for more details.
.
├── assets
│ ├── poses-viz.png
│ └── teaser.png
├── dataset
│ └── pi
├── env.yaml
├── log
├── README.md
├── snapshot
│ └── model-iter-40000.pth
├── src
│ ├── datasets
│ │ ├── data_utils_pi3d.py
│ │ ├── pi3d_hier.py
│ │ ├── pi3d.py
│ │ └── vis_2p.py
│ ├── model.py
│ ├── test.py
│ ├── train.py
│ └── utils
│ ├── angle_to_joint.py
│ ├── config.py
│ ├── logger.py
│ ├── misc.py
│ ├── parser.py
│ ├── pyt_utils.py
│ ├── rigid_align.py
│ ├── util.py
│ └── visualize.py
└── viz
conda env create -f env.yaml
conda activate bp42b
Request ExPI dataset here and place the pi folder under datasets/.
PYTHONPATH=. python src/train.py
PYTHONPATH=. python src/test.py
PYTHONPATH=. python src/test.py --visualize
On the common action split of ExPI dataset, we achieve the following results:
| 5 | 10 | 15 | 25 | |
|---|---|---|---|---|
| MPJPE | 40 | 87.1 | 130.1 | 201.3 |
| AME | 25 | 53 | 76 | 110 |
On the unseen action split of ExPI dataset, we achieve the following results:
| 10 | 15 | 20 | |
|---|---|---|---|
| MPJPE | 110.4 | 161.7 | 205.3 |
| AME | 65 | 93 | 114 |
We build upon siMLPe and take some code from MultiMotion.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent