Multi-Scale Adaptive Graph Convolutional Network for Skeleton-Based Action Recognition

Abstract

Skeleton-based action recognition is a branch of action recognition which uses dynamic skeletons as input. Recent research based on graph convolutional networks (GCN) has achieved remarkable performance in this area. However, feature extraction and fusion at different physical scales have not been well studied. To solve these issues, we propose a novel MultiScale Adaptive Graph Convolutional Network (MSGCN) which contains a Multi-Scale Graph Convolutional Module and a MultiScale Selective Fusion Module. Extensive experiments on NTU-RGBD dataset demonstrate the effectiveness of our method, our method achieved competitive performance on NTU-RGBD dataset.