Adaptive multi-level graph convolution with contrastive learning for skeleton-based action recognition

Graph Convolutional Networks (GCNs) have been widely used in skeleton-based action recognition with remarkable achievements. Many recent studies model the human body as a topological graph and extract action features using GCNs, however, the inherent shared and static characteristics of topological graph during training limit the performance of the model. In this work, we propose an adaptive multi-level graph convolution network (AML-GCN), which uses two different levels of spatial convolution to extract spatial features and enhances the feature representation by contrastive learning, then uses multiscale temporal convolution to capture advanced temporal features. We first propose dynamic topological spatial graph convolution (DTSGC) using dynamic graph topology, which can be continuously updated during training and has a strong generalization ability. To further enhance flexibility of the network, we present adaptive spatial graph convolution (ASGC) using non-shared graph topology, which is unique in different layers and can extract more diverse features. Meanwhile, we introduce contrastive learning to maximize mutual information between the two modules, which are trained together to promote each other and enhance feature representation. Extensive experiments on the three large datasets verify that the proposed model achieves excellent recognition accuracy compared to most current models.