Standard Deviation Labeling A New Partitioning Strategy for Spatial Temporal Graph Convolutional Networks in Human Interaction Recognition

抄録

Inter-body and intra-body connections of joints in human skeleton data are crucial information for graph-based models to perform better in human interaction recognition tasks. In this paper, we introduce a new partitioning strategy to label edges in skeleton graphs for the Spatial Temporal Graph Convolutional Networks (ST-GCN). This strategy is calculated based on the standard deviations of joint distances of two skeletons. Combining our strategy with the two-person graph and symmetry processing proposed in [1], we achieve better performance of the ST-GCN model on the human interaction recognition task. The experiments were conducted in four large and popular datasets: NTU RGB-D x-sub mutual, NTU RGB-D x-view mutual, NTU RGB-D120 x-sub mutual, and NTU RGB-D120 x-set mutual.