Graph Data Augmentation based on Adaptive Graph Convolution for Skeleton-based Action Recognition

Abstract

Recently, action recognition based on graph data has received widespread attention. Graph convolutional neural networks need a large amount of graph data support. Therefore, graph data augmentation has significant research value. Existing graph data augmentation methods are generally based on graph structure or graph node features to add or remove graph edges, leading to problems in generality and effectiveness. To improve the above issues, this paper utilizes adaptive graph convolution to design an algorithm that obtains the optimal graph connection for graph data augmentation as graph convolution proceeds, which greatly improves the generalization and effectiveness. The results on NTU-RGBD and Kinetics-Skeleton datasets for skeleton-based action recognition prove that our proposed method reveals better results than existing methods.