Zero-Shot-Action-Recognition-with-Two-Stream-GCN

This is a TensorFlow implementation of Two-Stream Graph Convolutional Networks for the task of action recognition, as described in our paper:

Junyu Gao, Tianzhu Zhang, Changsheng Xu, I Know the Relationships: Zero-Shot Action Recognition via Two-Stream Graph Convolutional Networks and Knowledge Graphs (AAAI 2019)

The code is developed based on the TensorFlow framework and the Graph Convolutional Network (GCN) repo GCN and zero-shot-gcn.

Prerequisite

• Construct a python enviorment with python3.6
• Install TensorFlow >= 1.2.0
  • Note the ZSAR performance slightly varies in different versions.
  • pip install tensorflow (For CPU)
  • pip install tensorflow-gpu (For GPU)
• Install networkx
  • pip install networkx

Dataset Preparation

The processed data of UCF101 can be downloaded from Google Driver. After downloading the file, unzip it to the folder ./

Run the demo

python train_two_stream_gcn.py

The above code will automatically train the two-stream GCN model on UCF101 dataset. The test accuracies will be outputted for each training epoch.

Cite

Please cite our paper if you use this code in your own work:

@inproceedings{junyu2019AAAI_TS-GCN,
  title={I Know the Relationships: Zero-Shot Action Recognition via Two-Stream Graph Convolutional Networks and Knowledge Graphs},
  author={Gao, Junyu and Zhang, Tianzhu and Xu,  Changsheng},
  booktitle={AAAI},
  year={2019}
}