This repository contains the code for the paper "PREDICT & CLUSTER: Unsupervised Skeleton Based Action Recognition", which is available here, published in CVPR 2020. The unsupervised approach implements two strategies discussed in the paper, Fixed-state(FS) and Fixed-weight(FW). It is tested on NW-UCLA and NTU-RGBD(60) dataset.

We propose a novel system for unsupervised skeleton-based action recognition. Given inputs of body-keypoints sequences obtained during various movements, our system associates the sequences with actions. Our system is based on an encoder-decoder recurrent neural network, where the encoder learns a separable feature representation within its hidden states formed by training the model to perform the prediction task. We show that according to such unsupervised training, the decoder and the encoder self-organize their hidden states into a feature space which clusters similar movements into the same cluster and distinct movements into distant clusters.

The python notebook ucla_demo.ipynb is the demonstration notebook for FS and FW strategies on NW-UCLA dataset, which is available on http://wangjiangb.github.io/my_data.html. The preprocessed UCLA data is included in ucla_data directory, please refer to ucla_demo.ipynb for more info. The data preprocessing part for UCLA dataset is also incorporated within the notebook, and is ready to run. The results shown in the notebook is not the most optimal one as we got in the paper, but is almost comparable.

Other python scripts are extracted from the notebooks we wrote for NTU-RGBD 60 datasets, but not tested yet.

1. Tensorflow 1.14.0
2. Python 3
3. scikit-learn 0.21.2
4. matplotlib 3.1.0
5. numpy 1.16.4

PyTorch Implementation (by Jingyuan Li @christincha) is avaialable. PyTorch implementation is part of extension of Predict-Cluster, Iterate-Cluster, to consider self-supervised active learning levarging Predict-Cluster. You can find more details about it here: https://arxiv.org/abs/2006.06911. Iterate-Cluster code will be available soon in a new linked github repository.

If you are using PyTorch, you may just need the ucla_github_pytorch folder. Please check the UCLA_Pytorch_Demo.ipynb for more details. The data for Pytorch implementation are preprocessed and saved as h5py files for PyTorch usage.

We share the preprocessed NTU data in Google Drive. The raw_train/test_data.pkl are cleaned data by removing the noise & multiple actors situations. The trans_train/test_data.pkl are data after applying view invriant transform. Check more details in the example script.

Since the original link of UWA3D dataset is down, I upload the raw data and the View invariant transformed data to Google Drive. Hope it helps.

Please cite the PREDICT & CLUSTER: Unsupervised Skeleton Based Action Recognition (CVPR 2020) when you use this code:

@inproceedings{su2020predict,
  title={Predict \& cluster: Unsupervised skeleton based action recognition},
  author={Su, Kun and Liu, Xiulong and Shlizerman, Eli},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={9631--9640},
  year={2020}
}