This is an official PyTorch implementation of "Zero-shot Skeleton-based Action Recognition via Mutual Information Estimation and Maximization" in ACMMM 2023.

[Paper]

# Install the python libraries
$ cd SMIE
$ pip install -r requirements.txt

We apply the same dataset processing as AimCLR.
You can also download the skeleton data in BaiduYun link:

• NTU-RGB-D 60.
• NTU-RGB-D 120.
• PKU-MMD.

The code: pstl

For the Semantic Features, You can download in BaiduYun link: Semantic Feature.

The code: smie

• [dataset]_embeddings.npy: based on label names using Sentence-Bert.
• [dataset]_clip_embeddings.npy: based on label names using CLIP.
• [dataset]_des_embeddings.npy: based on label descriptions using Sentence-Bert.

Put the semantic feautures in fold: ./data/language/

Using ChatGPT to expand each action label name into a complete action description. The total label descriptions can be found in folder.

Our SMIE employs two experiment setting.

• SynSE Experiment Setting: two datasets are used, split_5 and split_12 on NTU60, and split_10 and split_24 on NTU120. The visual feature extractor is Shift-GCN.
• Optimized Experiment Setting: three datasets are used (NTU-60, NTU-120, PKU-MMD), and each dataset have three random splits. The visual feature extractor is classical ST-GCN to minimize the impact of the feature extractor and focus on the connection model.

To compared with the SOTA method SynSE, we first apply their zero-shot class splits for SynSE Experiment Setting. You can download the visual features from their repo, or download from our BaiduYun link: SOTA visual features. Code:smie.

Example for training and testing on NTU-60 split_5 data.

# SynSE Experiment Setting
$ python procedure.py with 'train_mode="sota"'

You can also choose different split id of config.py (sota compare part).

For different class splits, you can change the split_id in split.py. Then run the split.py to obtain split data for different seen and unseen classes.

# class-split
$ python split.py

Refer to Generate_Feature.

Example for training and testing on NTU-60 split_1.
You can change some settings of config.py.

# Optimized Experiment Setting
$ python procedure.py with 'train_mode="main"'

If you find our paper and repo useful, please cite our paper. Thanks!

@inproceedings{zhou2023zero,
  title={Zero-shot Skeleton-based Action Recognition via Mutual Information Estimation and Maximization},
  author={Zhou, Yujie and Qiang, Wenwen and Rao, Anyi and Lin, Ning and Su, Bing and Wang, Jiaqi},
  booktitle={Proceedings of the 31st ACM International Conference on Multimedia},
  pages={5302--5310},
  year={2023}
}

• The codebase is from MS2L.
• The visual feature is based on ST-GCN.
• The semantic feature is based on Sentence-Bert.
• The baseline methods are from SynSE.

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

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