Multi-level Temporal-Guided Graph Convolutional Networks for Skeleton-Based Action Recognition

Abstract

Skeleton-based action recognition is a crucial and challenging task, which has promoted remarkable progress in diverse fields. Nevertheless, how to capture long-range temporal relationships remains a challenging problem, which is vital to reducing the ambiguity of indistinguishable actions. Towards this end, we propose a novel Multi-Level Temporal-Guided Graph Convolutional Network (ML-TGCN) to tackle the above problem. We leverage the multi-level temporal-guided mechanism to learn diverse temporal receptive fields for mining the discriminative motion patterns. Moreover, most current approaches cannot effectively explore the comprehensive spatial topology due to the skeleton graph is heuristically predefined, thus we propose a cross-space GCN to capture global context and maintain strengths of GCNs (i.e., hierarchy and local topology) jointly beyond the physical connectivity. The experimental results on the challenging datasets NTU RGB+D and Kinetics-Skeleton verify that ML-TGCN can achieve state-of-the-art performance.