A Graph Attention Spatio-temporal Convolutional Networks for 3D Human Pose Estimation in Video (GAST-Net)

News

[2020/08/14] We achieve real-time 3D pose estimation. [video]
[2020/10/15] We achieve real-time online 3D skeleton-based action recognition with a single RGB camera. [video]
[2020/11/17] We provide a tutorial on how to generate 3D poses/animation from a custom video. [INFERENCE_EN.md]

Introduction

Spatio-temporal information is key to resolve occlusion and depth ambiguity in 3D pose estimation. Previous methods have focused on either temporal contexts or local-to-global architectures that embed fixed-length spatio-temporal information. To date, there have not been effective proposals to simultaneously and flexibly capture varying spatio-temporal sequences and effectively achieves real-time 3D pose estimation. In this work, we improve the learning of kinematic constraints in the human skeleton: posture, local kinematic connections, and symmetry by modeling local and global spatial information via attention mechanisms. To adapt to single- and multi-frame estimation, the dilated temporal model is employed to process varying skeleton sequences. Also, importantly, we carefully design the interleaving of spatial semantics with temporal dependencies to achieve a synergistic effect. To this end, we propose a simple yet effective graph attention spatio-temporal convolutional network (GAST-Net) that comprises of interleaved temporal convolutional and graph attention blocks. Combined with the proposed method, we introduce a real-time strategy for online 3D skeleton-based action recognition with a simple RGB camera. Experiments on two challenging benchmark datasets (Human3.6M and HumanEva-I) and YouTube videos demonstrate that our approach effectively mitigates depth ambiguity and self-occlusion, generalizes to half upper body estimation, and achieves competitive performance on 2D-to-3D video pose estimation.

FrameWork

Real-time estimation

Dependencies

Make sure you have the following dependencies installed before proceeding:

Python >=3.6
PyTorch >= 1.0.1
matplotlib
numpy

Data preparation

Download the raw data from Human3.6M and HumanEva-I
Preprocess the dadaset in the same way as like VideoPose3D

Then put the preprocessed dataset under the data directory

 -data\
      data_2d_h36m_gt.npz
      data_3d_36m.npz
      data_2d_h36m_cpn_ft_h36m_dbb.npz
      data_2d_h36m_sh_ft_h36m.npz
  
      data_2d_humaneva15_gt.npz
      data_3d_humaneva15.npz
      data_2d_humaneva15_detectron_pt_coco.npz

Training & Testing

If you want to reproduce the results of our paper, run the following commands.

For Human3.6M:

python trainval.py -e 60 -k cpn_ft_h36m_dbb -arc 3,3,3 -drop 0.05 -b 128

For HumanEva:

python trainval.py -d humaneva15 -e 200 -k detectron_pt_coco -d humaneva15 -arc 3,3,3 -drop 0.5 -b 32 -lr 0.98 -str Train/S1,Train/S2,Train/S3 -ste Validate/S1,Validate/S2,Validate/S3 -a Walk,Jog,Box --by-subject

To test on Human3.6M, run:

python trainval.py -k cpn_ft_h36m_dbb -arc 3,3,3 -c checkpoint --evaluate epoch_60.bin

To test on HumanEva, run:

python trainval.py -k detectron_pt_coco -arc 3,3,3 -str Train/S1,Train/S2,Train/S3 -ste Validate/S1,Validate/S2,Validate/S3 -a Walk,Jog,Box --by-subject -c checkpoint --evaluate epoch_60.bin

Download our pretrained models

cd root_path
mkdir checkpoint output
cd checkpoint
mkdir gastnet

-checkpoint\gastnet\
            27_frame_model.bin
            81_frame_model.bin

Google Drive:

27 receptive field model

81 receptive field model

Baidu Yun(Extract code: kuxf):

27 receptive field model

81 receptive field model

Reconstruct 3D poses from 2D keypoints

Reconstruct 3D poses from 2D keypoints estimated from 2D detector (Mask RCNN, HRNet and OpenPose et al), and visualize it.

If you want to reproduce the baseball example, please run the following code:

python reconstruction.py

or run more detailed parameter settings:

python reconstruction.py -f 27 -w 27_frame_model.bin -k ./data/keypoints/baseball.json -vi ./data/video/baseball.mp4 -vo ./output/baseball_reconstruction.mp4 -kf coco

Reconstructed from YouTube video
Reconstructed from NTU-RGBD dataset

Inference a custom video

We provide a tutorial on how to run our model on custom videos. See INFERENCE.md for more details.

Acknowledgements

This repo is based on

Thanks to the original authors for their work!

Reference

If you find our paper and repo useful, please cite our paper. Thanks!

@article{liu2020a,
  title={A Graph Attention Spatio-temporal Convolutional Networks for 3D Human Pose Estimation in Video},
  author={Liu, Junfa and Rojas, Juan and Liang, Zhijun and Li, Yihui and Guan},
  journal={arXiv preprint arXiv:2003.14179},
  year={2020}
}

Contact

If you have any questions, please fell free to contact us. (junfaliu2019@gmail.com)

koustav123/GAST-Net-3DPoseEstimation