Skeleton‐based action recognition with JRR‐GCN

Abstract

A novel joints relation-reasoning, graph convolutional network (JRR-GCN) is proposed to solve the problem of skeleton-based action recognition (SAR). Different from the conventional spatial convolutional network-based methods, the adjacency matrices of JRR-GCN is reasoned by joints relation-reasoning network (JRR) automatically, which results in generating a more realistic representation of skeleton topology and yields better adjacency matrices for every sample. JRR is trained with the reinforcement learning with a novel state-action mapping scheme. Extensive experiments are conducted on two public SAR datasets, NTU-RGB+D and kinetics. Also the obtained results demonstrate the effectiveness of JRR-GCN comparing with the state-of-the-art SAR methods.