Abstract

Spatial-temporal graph convolutional networks (ST-GCNs) have been successfully applied for dynamic graphs representation learning, such as modeling skeleton-based human actions. However, ST-GCNs embed these non-Euclidean graph structures into Euclidean space, which is not the natural space to represent such structures as embedding them in this space incurs a large distortion. In this work, we make use of hyperbolic non-Euclidean geometry and construct compact ST-GCNs in the hyperbolic space. It can be shown that hyperbolic ST-GCNs (HST-GCNs) outperform the corresponding Euclidean counterparts. Additionally, these compact hyperbolic models can be used to increase the performance of large complex Euclidean models. Moreover, we show that the same or even better performance of large Euclidean models can be achieved by fusing the scores of smaller Euclidean models and a compact hyperbolic model. This in turn leads to reducing the total number of model parameters and hence model size. To validate the performance of these hyperbolic networks, we conducted extensive experiments on NTU RGB+D, NTU RGB+D 120 and Kinectics-Skeleton datasets for human action recognition.