Awesome

STST

STST: Spatial-Temporal Specialized Transformer for Skeleton-based Action Recognition in ACM MultiMedia2021

Result

A little different with paper due the reimplementation.

• NTU-RGB+D X-Sub: ~91.9%
• SHREC-28: ~95.3%

Packages Required

Python=3.6, Torch=1.6, Pickle, Numpy, Tqdm, Time, Opencv, Collections, Pyyaml, EasyDict, Shutil, Colorama, Argparse, TensorboardX, Itertools, Math, Inspect, Imutils

Data Preparation

• NTU-RGBD
  • Download the NTU-RGBD from the "https://rose1.ntu.edu.sg/dataset/actionRecognition"
  • Then you can use utils in folder gen_data to generate datasets for training.
• SHREC
  • We have provided and split the dataset here.

Training & Testing

• Train

  • Change the config file depending on what you want.

    python train.py --config ./config/shrec/shrec_stst_28.yaml

• Test

  • Change the config file depending on what you want.

    python eval.py --config ./workdir/val/shrec/stst_toy_28_val.yaml

    Here we provide a small version of the model that has been trained for you to test.

Citation

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

@inproceedings{zhang2021stst,
    title={STST: Spatial-Temporal Specialized Transformer for Skeleton-based Action Recognition},
    author={Zhang, Yuhan and Wu, Bo and Li, Wen and Duan, Lixin and Gan, Chuang},
    booktitle={Proceedings of the 29th ACM International Conference on Multimedia},
    pages={3229--3237},
    year={2021}
}

Reference

The code of this project is based on the DSTA-Net(Decoupled Spatial-Temporal Attention Network for Skeleton-Based Action-Gesture Recognition)

Contact

For any questions, feel free to contact: yuhanzhan9@gmail.com