Szy-Young/AGANet

Attention-Oriented Action Recognition for Real-Time Human-Robot Interaction (ICPR 2020)

AGANet (Attention-Guided Action Network)

Paper | Poster | ICPR page

This is the code implementation for AGANet, an attention-guided deep convolutional network for skeleton-based action recognition, proposed by our paper "Attention-Oriented Action Recognition for Real-Time Human-Robot Interaction" in ICPR 2020.

Requirements

tensorflow 1.9.0
numpy
matplotlib (for visualization only)

Core contents

datagen.py      # base Dataset
datagen_aug.py      # base Dataset + augmentation

cnn_model.py        # base CNN
cnn_model_att.py        # base CNN + attention

How to use

python train.py     # Training

python test_AP.py
python test_cAP.py      # Quantitative evaluation

python error_analysis.py        # Compute confusion matrix

Citation

@inproceedings{song2021attention,
  title={Attention-Oriented Action Recognition for Real-Time Human-Robot Interaction},
  author={Song, Ziyang and Yin, Ziyi and Yuan, Zejian and Zhang, Chong and Chi, Wanchao and Ling, Yonggen and Zhang, Shenghao},
  booktitle={2020 25th International Conference on Pattern Recognition (ICPR)},
  pages={7087--7094},
  year={2021},
  organization={IEEE}
}