Temporal-Aware Graph Convolution Network for Skeleton-based Action Recognition

Graph convolutions networks (GCN) have drawn attention for skeleton-based action recognition because a skeleton with joints and bones can be naturally regarded as a graph structure. However, the existing methods are limited in temporal sequence modeling of human actions. To consider temporal factors in action modeling, we present a novel Temporal-Aware Graph Convolution Network (TA-GCN). First, we design a causal temporal convolution (CTCN) layer to ensure no impractical future information leakage to the past. Second, we present a novel cross-spatial-temporal graph convolution (3D-GCN) layer that extends an adaptive graph from the spatial to the temporal domain to capture local cross-spatial-temporal dependencies among joints. Involving the two temporal factors, TA-GCN can model the sequential nature of human actions. Experimental results on two large-scale datasets, NTU-RGB+D and Kinetics-Skeleton, indicate that our network achieves accuracy improvement (about 1% on the two datasets) over previous methods.