An unofficial Tensorflow implementation of the paper "Spatial Temporal Graph Convolutional Networks for Skeleton-Based Action Recognition" in AAAI 2018.

• Paper: PDF
• Code is based on mmskeleton and 2s-AGCN

Model weights for ST-GCN trained on xview and xsub joint data Dropbox

• Python >= 3.5
• scipy >= 1.3.0
• numpy >= 1.16.4
• tensorflow >= 2.0.0

Most of the interesting stuff can be found in:

• model/stgcn.py: model definition of ST-GCN
• data_gen/: how raw datasets are processed into numpy tensors
• graphs/ntu_rgb_d.py: graph definition
• main.py: general training/eval processes; etc.

1. The NTU RGB+D dataset can be downloaded from here. We'll only need the Skeleton data (~ 5.8G).

2. After downloading, unzip it and put the folder nturgb+d_skeletons to ./data/nturgbd_raw/.

3. Generate the joint dataset first:

cd data_gen
python3 gen_joint_data.py

Specify the data location if the raw skeletons data are placed somewhere else. The default looks at ./data/nturgbd_raw/.

4. Generate the tfrecord files for joint data :

python3 gen_tfrecord_data.py

To start training the network with the joint data, use the following command:

python3 main.py --train-data-path data/ntu/<dataset folder> --test-data-path data/ntu/<dataset folder>

Here refers to the folder containing the tfrecord files generated in step 5 of the pre-processing steps.

Note: At the moment, only nturgbd-cross-subject is supported.

Please cite the following paper if you use this repository in your reseach

@inproceedings{yan2018spatial,
      title={Spatial temporal graph convolutional networks for skeleton-based action recognition},
      author={Yan, Sijie and Xiong, Yuanjun and Lin, Dahua},
      booktitle={Thirty-Second AAAI Conference on Artificial Intelligence},
      year={2018}
}