Reproduce “Topology-aware Convolutional Neural Network for Efficient Skeleton-based Action Recognition” on OpenResearch

About this paper

Kailin Xu, Fanfan Ye, Qiaoyong Zhong, Di Xie

Researchers from Zhejiang University and Hikvision Research Institute developed Topology-aware CNN (Ta-CNN), an efficient pure CNN architecture for skeleton-based action recognition. The Ta-CNN achieves competitive performance with state-of-the-art Graph Convolutional Networks while demonstrating significantly lower computational costs, and it introduces a specialized data augmentation method called SkeletonMix.

Key insights

• Pure CNNs can effectively model the irregular topological structure of skeleton data and achieve competitive performance with GCNs when specifically designed with topology-aware mechanisms.
• Graph convolution can be understood as a specific instance of normal convolution, suggesting a deeper theoretical connection and shared capabilities between the two architectures for graph-structured data.
• Explicitly enhancing coordinate and joint features through modules like CAG and VAG allows CNNs to capture fine-grained distinctions in human actions, improving recognition, especially for challenging, subtle movements.

Summary by alphaXiv.