chrischen1023 / cp-agcn

CP-AGCN: A Pytorch-based Attention Informed Graph Convolutional Networks for Cerebral Palsy Classification

The early diagnosis is clinically considered one of the essential parts of cerebral palsy (CP) treatment, so we propose to design a low-cost and interpretable classification system for supporting CP diagnosis. In this work, we implement a Pytorch-based attention-informed graph convolutional network to classify CP patients. This is achieved by integrating the additive attention mechanism into the graph convolutional network. We also propose an optional frequency-binning module to learn the CP movements in the frequency domain while filtering noise. The current version system only requires consumer-grade RGB videos for training to support interactive-time CP diagnosis by providing an interpretable CP classification result. Our flexible system can be further extended to handle other human motion-related disorders (e.g., freezing of gait) and human action recognition tasks.

For the full MINI-RGBD dataset, please refer to https://www.iosb.fraunhofer.de/en/competences/image-exploitation/object-recognition/sensor-networks/motion-analysis.html

For the full RVI-38 dataset, please refer to https://github.com/edmondslho/Pose-basedCerebralPalsyPrediction

Initialization

python >= 3.7

pytorch

Library requirement and installation

pip install requirement.txt cd torchlight; python setup.py install; cd ..

Getting started

The command of a quick training with Leave-One-Out Cross-Validation on the MINI-RGBD dataset.

python start.py