Structural Topology Refinement Network for Skeleton-Based Action Recognition

The effectiveness of graph convolutional networks (GCNs) has been widely demonstrated in skeleton-based action recognition. However, most existing GCN-based methods use a dense adjacency matrix to describe the structural information of the entire skeleton, i.e., a holistic representation, which neglects the discriminability of local patterns. To address this challenge, we propose a novel region descriptor by dividing the skeleton into different local sections (i.e., left arm, right arm, left leg, right leg, torso, and head). The generated representations contain rich semantic information, enabling the model to better understand the action correlation between different body parts. Inspired by the success of manifold learning in nonlinear data characterization, the symmetric positive definite (SPD) matrix and Riemannian neural network are further introduced to capture the long-range statistical relationships among different topographies. These components form our structural topology refi...

The effectiveness of graph convolutional networks (GCNs) has been widely demonstrated in skeleton-based action recognition. However, most existing GCN-based methods use a dense adjacency matrix to describe the structural information of the entire skeleton, i.e., a holistic representation, which neglects the discriminability of local patterns. To address this challenge, we propose a novel region descriptor by dividing the skeleton into different local sections (i.e., left arm, right arm, left leg, right leg, torso, and head). The generated representations contain rich semantic information, enabling the model to better understand the action correlation between different body parts. Inspired by the success of manifold learning in nonlinear data characterization, the symmetric positive definite (SPD) matrix and Riemannian neural network are further introduced to capture the long-range statistical relationships among different topographies. These components form our structural topology refinement network (STRN). Extensive experiments on three benchmark datasets, namely NTU-60, NTU-120, and NW-UCLA, show the superiority of our proposed method over the state-of-the-art (SOTA).