Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/Walter0807/MotionBERT
[ICCV 2023] PyTorch Implementation of "MotionBERT: A Unified Perspective on Learning Human Motion Representations"
https://github.com/Walter0807/MotionBERT
3d-pose-estimation iccv2023 mesh-recovery skeleton-based-action-recognition
Last synced: about 1 month ago
JSON representation
[ICCV 2023] PyTorch Implementation of "MotionBERT: A Unified Perspective on Learning Human Motion Representations"
- Host: GitHub
- URL: https://github.com/Walter0807/MotionBERT
- Owner: Walter0807
- License: apache-2.0
- Created: 2022-05-02T13:29:37.000Z (over 2 years ago)
- Default Branch: main
- Last Pushed: 2024-06-18T16:01:46.000Z (6 months ago)
- Last Synced: 2024-08-03T04:06:12.698Z (5 months ago)
- Topics: 3d-pose-estimation, iccv2023, mesh-recovery, skeleton-based-action-recognition
- Language: Python
- Homepage:
- Size: 103 KB
- Stars: 967
- Watchers: 22
- Forks: 118
- Open Issues: 36
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# MotionBERT: A Unified Perspective on Learning Human Motion Representations
[](https://arxiv.org/abs/2210.06551) 
[](https://huggingface.co/walterzhu/MotionBERT)
[](https://paperswithcode.com/sota/monocular-3d-human-pose-estimation-on-human3?p=motionbert-unified-pretraining-for-human)
[](https://paperswithcode.com/sota/one-shot-3d-action-recognition-on-ntu-rgbd?p=motionbert-unified-pretraining-for-human)
[](https://paperswithcode.com/sota/3d-human-pose-estimation-on-3dpw?p=motionbert-unified-pretraining-for-human)
This is the official PyTorch implementation of the paper *"[MotionBERT: A Unified Perspective on Learning Human Motion Representations](https://arxiv.org/pdf/2210.06551.pdf)"* (ICCV 2023).
## Installation
```bash
conda create -n motionbert python=3.7 anaconda
conda activate motionbert
# Please install PyTorch according to your CUDA version.
conda install pytorch torchvision torchaudio pytorch-cuda=11.6 -c pytorch -c nvidia
pip install -r requirements.txt
```
## Getting Started
| Task | Document |
| --------------------------------- | ------------------------------------------------------------ |
| Pretrain | [docs/pretrain.md](docs/pretrain.md) |
| 3D human pose estimation | [docs/pose3d.md](docs/pose3d.md) |
| Skeleton-based action recognition | [docs/action.md](docs/action.md) |
| Mesh recovery | [docs/mesh.md](docs/mesh.md) |
## Applications
### In-the-wild inference (for custom videos)
Please refer to [docs/inference.md](docs/inference.md).
### Using MotionBERT for *human-centric* video representations
```python
'''
x: 2D skeletons
type =
shape = [batch size * frames * joints(17) * channels(3)]
MotionBERT: pretrained human motion encoder
type =
E: encoded motion representation
type =
shape = [batch size * frames * joints(17) * channels(512)]
'''
E = MotionBERT.get_representation(x)
```
> **Hints**
>
> 1. The model could handle different input lengths (no more than 243 frames). No need to explicitly specify the input length elsewhere.
> 2. The model uses 17 body keypoints ([H36M format](https://github.com/JimmySuen/integral-human-pose/blob/master/pytorch_projects/common_pytorch/dataset/hm36.py#L32)). If you are using other formats, please convert them before feeding to MotionBERT.
> 3. Please refer to [model_action.py](lib/model/model_action.py) and [model_mesh.py](lib/model/model_mesh.py) for examples of (easily) adapting MotionBERT to different downstream tasks.
> 4. For RGB videos, you need to extract 2D poses ([inference.md](docs/inference.md)), convert the keypoint format ([dataset_wild.py](lib/data/dataset_wild.py)), and then feed to MotionBERT ([infer_wild.py](infer_wild.py)).
>
## Model Zoo
| Model | Download Link | Config | Performance |
| ------------------------------- | ------------------------------------------------------------ | ------------------------------------------------------------ | ---------------- |
| MotionBERT (162MB) | [OneDrive](https://1drv.ms/f/s!AvAdh0LSjEOlgS425shtVi9e5reN?e=6UeBa2) | [pretrain/MB_pretrain.yaml](configs/pretrain/MB_pretrain.yaml) | - |
| MotionBERT-Lite (61MB) | [OneDrive](https://1drv.ms/f/s!AvAdh0LSjEOlgS27Ydcbpxlkl0ng?e=rq2Btn) | [pretrain/MB_lite.yaml](configs/pretrain/MB_lite.yaml) | - |
| 3D Pose (H36M-SH, scratch) | [OneDrive](https://1drv.ms/f/s!AvAdh0LSjEOlgSvNejMQ0OHxMGZC?e=KcwBk1) | [pose3d/MB_train_h36m.yaml](configs/pose3d/MB_train_h36m.yaml) | 39.2mm (MPJPE) |
| 3D Pose (H36M-SH, ft) | [OneDrive](https://1drv.ms/f/s!AvAdh0LSjEOlgSoTqtyR5Zsgi8_Z?e=rn4VJf) | [pose3d/MB_ft_h36m.yaml](configs/pose3d/MB_ft_h36m.yaml) | 37.2mm (MPJPE) |
| Action Recognition (x-sub, ft) | [OneDrive](https://1drv.ms/f/s!AvAdh0LSjEOlgTX23yT_NO7RiZz-?e=nX6w2j) | [action/MB_ft_NTU60_xsub.yaml](configs/action/MB_ft_NTU60_xsub.yaml) | 97.2% (Top1 Acc) |
| Action Recognition (x-view, ft) | [OneDrive](https://1drv.ms/f/s!AvAdh0LSjEOlgTaNiXw2Nal-g37M?e=lSkE4T) | [action/MB_ft_NTU60_xview.yaml](configs/action/MB_ft_NTU60_xview.yaml) | 93.0% (Top1 Acc) |
| Mesh (with 3DPW, ft) | [OneDrive](https://1drv.ms/f/s!AvAdh0LSjEOlgTmgYNslCDWMNQi9?e=WjcB1F) | [mesh/MB_ft_pw3d.yaml](configs/mesh/MB_ft_pw3d.yaml) | 88.1mm (MPVE) |
In most use cases (especially with finetuning), `MotionBERT-Lite` gives a similar performance with lower computation overhead.
## TODO
- [x] Scripts and docs for pretraining
- [x] Demo for custom videos
## Citation
If you find our work useful for your project, please consider citing the paper:
```bibtex
@inproceedings{motionbert2022,
title = {MotionBERT: A Unified Perspective on Learning Human Motion Representations},
author = {Zhu, Wentao and Ma, Xiaoxuan and Liu, Zhaoyang and Liu, Libin and Wu, Wayne and Wang, Yizhou},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision},
year = {2023},
}
```