dynamic_skeleton

Pytorch implementation of paper '' A Module Selection-based Approach for Efficient Skeleton Human Action Recognition''. Our method can be easily combined to other state-of-the-art skeleton-based backbone networks, we provide CTR-GCN as backbone in this repo.

Data Preparation

Please follow the instructions of CTR-GCN.

Training & Testing

Training

# Example: training network on NTU RGB+D 60 cross subject
python main.py --config ./config/nturgbd-cross-subject/default.yaml

Testing

# Example: testing the joint modality of nturgbd-cross-subject dataset using second scheme
python main.py --config ./config/nturgbd-cross-subject/default.yaml --phase test --save-score True --weights weight/CTR-GCN-Scheme2/ntu60/xsub/CS_joint.pt --model model.dynamic_ctrgcn_scheme2_test.Model

• To ensemble the results of different modalities, run:

# Example: ensemble four modalities on NTU RGB+D 60 cross subject
python ensemble.py --datasets ntu/xsub --joint-dir work_dir/ntu/xsub/ctrgcn --bone-dir work_dir/ntu/xsub/ctrgcn_bone --joint-motion-dir work_dir/ntu/xsub/ctrgcn_motion --bone-motion-dir work_dir/ntu/xsub/ctrgcn_bone_motion

Pretrained Models

• Pretrained Models are avaliable at https://drive.google.com/file/d/16tBLRYZszba8-nK1GVfQ-ESHVuM-c2yG/view?usp=share_link.

Acknowledgements

This repo is based on CTR-GCN.

Thanks original authors for their work!