This is an official PyTorch implementation of "Contrastive Learning from Extremely Augmented Skeleton Sequences for Self-supervised Action Recognition" in AAAI2022.

• Download the raw data of NTU RGB+D and PKU-MMD.
• For NTU RGB+D dataset, preprocess data with tools/ntu_gendata.py. For PKU-MMD dataset, preprocess data with tools/pku_part1_gendata.py.
• Then downsample the data to 50 frames with feeder/preprocess_ntu.py and feeder/preprocess_pku.py.
• If you don't want to process the original data, download the file folder in Google Drive action_dataset or BaiduYun link action_dataset, code: 0211. NTU-120 is also provided: NTU-120-frame50.

# Install torchlight
$ cd torchlight
$ python setup.py install
$ cd ..

# Install other python libraries
$ pip install -r requirements.txt

Example for unsupervised pre-training of 3s-AimCLR. You can change some settings of .yaml files in config/ntu60/pretext folder.

# train on NTU RGB+D xview joint stream
$ python main.py pretrain_aimclr --config config/ntu60/pretext/pretext_aimclr_xview_joint.yaml

# train on NTU RGB+D xview motion stream
$ python main.py pretrain_aimclr --config config/ntu60/pretext/pretext_aimclr_xview_motion.yaml

# train on NTU RGB+D xview bone stream
$ python main.py pretrain_aimclr --config config/ntu60/pretext/pretext_aimclr_xview_bone.yaml

Example for linear evaluation of 3s-AimCLR. You can change .yaml files in config/ntu60/linear_eval folder.

# Linear_eval on NTU RGB+D xview
$ python main.py linear_evaluation --config config/ntu60/linear_eval/linear_eval_aimclr_xview_joint.yaml

$ python main.py linear_evaluation --config config/ntu60/linear_eval/linear_eval_aimclr_xview_motion.yaml

$ python main.py linear_evaluation --config config/ntu60/linear_eval/linear_eval_aimclr_xview_bone.yaml

We release several trained models in released_model. The performance is better than that reported in the paper. You can download them and test them with linear evaluation by changing weights in .yaml files.

For three-streams results, you can train three separate models and ensemble the results, or you can use three models in one .py file, similar to net/crossclr_3views.py.

The t-SNE visualization of the embeddings after AimCLR pre-training on NTU60-xsub.

Please cite our paper if you find this repository useful in your resesarch:

@inproceedings{guo2022aimclr,
  Title= {Contrastive Learning from Extremely Augmented Skeleton Sequences for Self-supervised Action Recognition},
  Author= {Tianyu, Guo and Hong, Liu and Zhan, Chen and Mengyuan, Liu and Tao, Wang  and Runwei, Ding},
  Booktitle= {AAAI},
  Year= {2022}
}

The framework of our code is extended from the following repositories. We sincerely thank the authors for releasing the codes.

• The framework of our code is based on CrosSCLR.
• The encoder is based on ST-GCN.

This project is licensed under the terms of the MIT license.