TD-GCN-Gesture

This is the official repo of TD-GCN and our work is accepted by IEEE Transactions on Multimedia (TMM).
Jinfu Liu, Xinshun Wang, Can Wang, Yuan Gao, Mengyuan Liu. Temporal Decoupling Graph Convolutional Network for Skeleton-based Gesture Recognition. IEEE Transactions on Multimedia (TMM), 2023.

Prerequisites

You can install all dependencies by running pip install -r requirements.txt
Then, you need to install torchlight by running pip install -e torchlight

Data Preparation

Download four datasets:

1. SHREC’17 Track dataset from http://www-rech.telecom-lille.fr/shrec2017-hand/
   
2. DHG-14/28 dataset from http://www-rech.telecom-lille.fr/DHGdataset/
   
3. NTU RGB+D 60 Skeleton dataset from https://rose1.ntu.edu.sg/dataset/actionRecognition/
   
4. NW-UCLA dataset from Download NW-UCLA dataset
   
5. Put downloaded data into the following directory structure:

- data/
  - shrec/
    - shrec17_dataset/
	  - HandGestureDataset_SHREC2017/
	    - gesture_1
	      ...
  - DHG14-28/
    - DHG14-28_dataset/
	  - gesture_1
	    ...
  - NW-UCLA/
    - all_sqe
      ...
  - ntu/
    - nturgbd_raw/
	  - nturgb+d_skeletons
            ...

SHREC’17 Track dataset:

1. First, extract all files to /data/shrec/shrec17_dataset
   
2. Then, run python gen_traindataset.py and python gen_testdataset.py
   

DHG-14/28 dataset:

1. First, extract all files to ./data/DHG14-28/DHG14-28_dataset
   
2. Then, run python python gen_dhgdataset.py

NTU RGB+D 60 dataset

1. First, extract all skeleton files to ./data/ntu/nturgbd_raw
   
2. Then, run python get_raw_skes_data.py, python get_raw_denoised_data.py and python seq_transformation.py in sequence
   

NW-UCLA dataset

1. Move folder all_sqe to ./data/NW-UCLA

Training

You can change the configuration in the yaml file and in the main function. We also provide four default yaml configuration files.

SHREC’17 Track dataset:

Run python main.py --device 0 --config ./config/shrec17/shrec17.yaml

DHG-14/28 dataset:

Run python main.py --device 0 --config ./config/dhg14-28/DHG14-28.yaml

NTU RGB+D 60 dataset:

On the benchmark of cross-view, run python main.py --device 0 --config ./config/nturgbd-cross-view/default.yaml
On the benchmark of cross-subject, run python main.py --device 0 --config ./config/nturgbd-cross-subject/default.yaml

NW-UCLA dataset:

Run python main.py --device 0 --config ./config/ucla/nw-ucla.yaml

Testing

We provide several trained weight files and place them in the checkpoints folder.

python main.py --device 0 --config <config.yaml> --phase test --weights <work_dir>/<weight.pt>

Ensemble

1. Set Rate
2. Run:
python gesture_ensemble.py \
--joint_Score <joint_path> \
--bone_Score <bone_path> \
--jointmotion_Score <jointmotion_path> \
--val_sample <val_path> \
--benchmark <benchmark>

# Example for Shrec_28
1. Download .pkl file from: https://drive.google.com/drive/folders/1ux87mUirBQjmA4b4fEWtb9tuj-8wSYHt
2. Set Rate [0.5, 0.5, 0.5] or [0.5, 0.3, 0.2]
3. Run:
python gesture_ensemble.py \
--joint_Score ./joint.pkl \
--bone_Score ./bone.pkl \
--jointmotion_Score ./jointmotion.pkl \
--val_sample ./shrec17_28.txt \
--benchmark Shrec_28

Citation

@ARTICLE{10113233,
  author={Liu, Jinfu and Wang, Xinshun and Wang, Can and Gao, Yuan and Liu, Mengyuan},
  journal={IEEE Transactions on Multimedia}, 
  title={Temporal Decoupling Graph Convolutional Network for Skeleton-based Gesture Recognition}, 
  year={2024}
}

Our project is based on the DSTA-Net, CTR-GCN.

Contact

For any questions, feel free to contact: [email protected]