Inception Spatial Temporal Graph Convolutional Networks for Skeleton-Based Action Recognition

Abstract

Graph convolutional networks is widely used in the field of skeleton-based motion recognition because of its characteristics of applying to non-Euclidean data. But most of the existing methods based on graph convolutional networks only perform convolution between adjacent joint points, ignoring the connection with farther joint points and symmetrical points. In order to improve the accuracy of skeleton-based action recognition, we propose a novel Inception spatial temporal graph convolutional networks (IST-GCN) for skeleton-based action recognition. By introducing the symmetry characteristics of the skeleton, our model can extract the interactive features of the symmetrical part of the human body. We also use the idea of multi-scale convolution to improve graph convolutional networks and temporal convolutional networks based on the Inception structure to better extract spatial and temporal features. A large number of experiments on NTU-RGB+D dataset show that our models have achieved higher accuracy and are more suitable for application in skeleton-based motion recognition. Codes are available at https://github.com/julycrow/IST-GCN.