Fully Distance-Guided Refinement Graph Convolutional Networks for Skeleton-Based Human Action Recognition

Graph convolutional networks (GCNs) have obtained impressive performance on skeleton-based human action recognition tasks in recent years. In GCNs, how to capture discriminative spatial temporal dependency features with graph modeling rules is the key for action recognition. However, previous GCNs usually relying on the heuristic spatial partition and unconstrained dynamic graph learning strategies may result in limited receptive field and misalignment with prior graph modeling rules. To solve the above issues, we propose a fully distance-guided refinement graph convolutional networks (FDR-GCN) to extract discriminative dependency features for increasing the performance and robustness of graph convolutions. In spatial graph modeling process, we construct a fully distance-guided graph containing all adjacent joint connections under different distance ranges to expand the spatial dependencies in one graph convolution operation. Refined graph convolution is then proposed to learn the distance-guided graph edge importance and keep it normalized simultaneously. In addition, we extend the bottleneck structure from temporal modeling to a unified spatial temporal graph modeling block, reducing the computational complexity with ratio up to 2.23× versus baseline method. Extensive experiments on three datasets, NTU-RGBD, NTU-RGBD 120 and NW-UCLA show that our proposed method surpasses previous state-of-the-art methods.