MotionBERT / docs /pretrain.md
walterzhu's picture
Upload 58 files
bbde80b
# Pretrain
## Data
### AMASS
1. Please download data from the [official website](https://amass.is.tue.mpg.de/download.php) (SMPL+H).
2. We provide the preprocessing scripts as follows. Minor modifications might be necessary.
- [tools/compress_amass.py](../tools/compress_amass.py): downsample the frame rate
- [tools/preprocess_amass.py](../tools/preprocess_amass.py): render the mocap data and extract the 3D keypoints
- [tools/convert_amass.py](../tools/convert_amass.py): slice them to motion clips
### Human 3.6M
Please refer to [pose3d.md](pose3d.md#data).
### InstaVariety
1. Please download data from [human_dynamics](https://github.com/akanazawa/human_dynamics/blob/master/doc/insta_variety.md#generating-tfrecords) to `data/motion2d`.
1. Use [tools/convert_insta.py](../tools/convert_insta.py) to preprocess the 2D keypoints (need to specify `name_action` ).
### PoseTrack
Please download PoseTrack18 from [MMPose](https://mmpose.readthedocs.io/en/latest/tasks/2d_body_keypoint.html#posetrack18) and unzip to `data/motion2d`.
The processed directory tree should look like this:
```
.
└── data/
β”œβ”€β”€ motion3d/
β”‚ └── MB3D_f243s81/
β”‚ β”œβ”€β”€ AMASS
β”‚ └── H36M-SH
β”œβ”€β”€ motion2d/
β”‚ β”œβ”€β”€ InstaVariety/
β”‚ β”‚ β”œβ”€β”€ motion_all.npy
β”‚ β”‚ └── id_all.npy
β”‚ └── posetrack18_annotations/
β”‚ β”œβ”€β”€ train
β”‚ └── ...
└── ...
```
## Train
```bash
python train.py \
--config configs/pretrain/MB_pretrain.yaml \
-c checkpoint/pretrain/MB_pretrain
```