LST

This repo is the official implementation for "LST: Language Supervised Training for Skeleton-based Action Recognition"

Introduction

Skeleton-based action recognition has drawn a lot of attention for its computation efficiency and robustness to lighting conditions. Existing skeleton-based action recognition methods are typically formulated as a one-hot classification task without fully utilizing the semantic relations between actions. For example, "make victory sign" and "thumb up" are two actions of hand gestures, whose major difference lies in the movement of hands. This information is agnostic from the categorical one-hot encoding of action classes but could be unveiled in the language description of actions. Therefore, utilizing action language descriptions in training could potentially benefit representation learning. In this work, we propose a Language Supervised Training (LST) approach for skeleton-based action recognition. More specifically, we employ a large-scale language model as the knowledge engine to provide text descriptions for body parts movements of actions, and propose a multi-modal training scheme by utilizing the text encoder to generate feature vectors for different body parts and supervise the skeleton encoder for action representation learning. Experiments show that our proposed LST method achieves noticeable improvements over various baseline models without extra computation cost at inference. LST achieves new state-of-the-arts on popular skeleton-based action recognition benchmarks, including NTU RGB+D, NTU RGB+D 120 and NW-UCLA.

Architecture of LST

Prerequisites

Python >= 3.6
PyTorch >= 1.1.0
PyYAML, tqdm, tensorboardX
We provide the dependency file of our experimental environment, you can install all dependencies by creating a new anaconda virtual environment and running pip install -r requirements.txt
Run pip install -e torchlight

Data Preparation

Please follow CTR-GCN for data preparation

Download datasets.

There are 3 datasets to download:

NTU RGB+D 60 Skeleton
NTU RGB+D 120 Skeleton
NW-UCLA

NTU RGB+D 60 and 120

Request dataset here: https://rose1.ntu.edu.sg/dataset/actionRecognition
Download the skeleton-only datasets:
1. nturgbd_skeletons_s001_to_s017.zip (NTU RGB+D 60)
2. nturgbd_skeletons_s018_to_s032.zip (NTU RGB+D 120)
3. Extract above files to ./data/nturgbd_raw

NW-UCLA

Download dataset from here
Move all_sqe to ./data/NW-UCLA

Data Processing

Directory Structure

Put downloaded data into the following directory structure:

- data/
  - NW-UCLA/
    - all_sqe
      ... # raw data of NW-UCLA
  - ntu/
  - ntu120/
  - nturgbd_raw/
    - nturgb+d_skeletons/     # from `nturgbd_skeletons_s001_to_s017.zip`
      ...
    - nturgb+d_skeletons120/  # from `nturgbd_skeletons_s018_to_s032.zip`
      ...

Generating Data

Generate NTU RGB+D 60 or NTU RGB+D 120 dataset:

 cd ./data/ntu # or cd ./data/ntu120
 # Get skeleton of each performer
 python get_raw_skes_data.py
 # Remove the bad skeleton 
 python get_raw_denoised_data.py
 # Transform the skeleton to the center of the first frame
 python seq_transformation.py

Training & Testing

Training

To train model on NTU60/120

# Example: training LST on NTU RGB+D cross subject joint modality
CUDA_VISIBLE_DEVICES=0,1 python main_multipart_ntu.py --config config/nturgbd-cross-subject/lst_joint.yaml --model model.ctrgcn.Model_lst_4part --work-dir work_dir/ntu60/csub/lst_joint --device 0 1
# Example: training LST on NTU RGB+D cross subject bone modality
CUDA_VISIBLE_DEVICES=0,1 python main_multipart_ntu.py --config config/nturgbd-cross-subject/lst_bone.yaml --model model.ctrgcn.Model_lst_4part_bone --work-dir work_dir/ntu60/csub/lst_bone --device 0 1
# Example: training LST on NTU RGB+D 120 cross subject joint modality
CUDA_VISIBLE_DEVICES=0,1 python main_multipart_ntu.py --config config/nturgbd120-cross-subject/lst_joint.yaml --model model.ctrgcn.Model_lst_4part --work-dir work_dir/ntu120/csub/lst_joint --device 0 1
# Example: training LST on NTU RGB+D 120 cross subject bone modality
CUDA_VISIBLE_DEVICES=0,1 python main_multipart_ntu.py --config config/nturgbd120-cross-subject/lst_bone.yaml --model model.ctrgcn.Model_lst_4part_bone --work-dir work_dir/ntu120/csub/lst_bone --device 0 1

To train model on NW-UCLA

CUDA_VISIBLE_DEVICES=0,1 python main_multipart_ucla.py --config config/ucla/lst_joint.yaml --model model.ctrgcn.Model_lst_4part_ucla --work-dir work_dir/ucla/lst_joint --device 0 1

Testing

To test the trained models saved in <work_dir>, run the following command:

python main_multipart_ntu.py --config <work_dir>/config.yaml --work-dir <work_dir> --phase test --save-score True --weights <work_dir>/xxx.pt --device 0

To ensemble the results of different modalities, run

# Example: ensemble four modalities of LST on NTU RGB+D 120 cross subject
python ensemble.py --datasets ntu120/xsub --joint-dir work_dir/ntu120/csub/lst_joint --bone-dir work_dir/ntu120/csub/lst_bone --joint-motion-dir work_dir/ntu120/csub/lst_joint_vel --bone-motion-dir work_dir/ntu120/csub/lst_bone_vel

Acknowledgements

This repo is based on CTR-GCN. The data processing is borrowed from SGN and HCN. The code for different modality is adopted from InfoGCN. The implementation for contrastive loss is adopted from ActionCLIP.

Thanks to the original authors for their work!

Citation

Please cite this work if you find it useful:

@article{xiang2022lst,
    title={Language Supervised Training for Skeleton-based Action Recognition},
    author={Wangmeng Xiang, Chao Li, Yuxuan Zhou, Biao Wang, Lei Zhang},
    journal={arXiv preprint arXiv:2208.05318},
    year={2022}
}

LiChenyang-Github / LST

LST