Differential Learning and Parallel Convolutional Network for Skeleton-Based Action Recognition

Abstract

Graph convolution network (GCN) has recently played a positive role in improving the accuracy of skeleton-based action recognition. Many GCN methods have reached a high accuracy. However, the lightweight of network model has recently become a major concern. Pointing at the problem, this paper introduces a lightweight network, a Differential Learning and Parallel Convolutional Network (DL-PCN), which is based on Semantics-Guided Neural Networks (SGN). The network is mainly composed of Differential Learning Module (DLM) and Parallel Convolutional Network Module (PCN). DLM is characterized by the feedforward connection, which improves the experiment accuracy. PCN can learn the multi-dimensional information of original skeleton data by the parallel connection of GCN and convolutional neural network (CNN). Considering the test accuracy of action recognition and network parameters, our network achieves the comparable performance on the NTU RGB+D 60 dataset and the NTU RGB+D 120 dataset.