Spatial-Temporal Graph Convolutional Networks for Action Recognition with Adjacency Matrix Generation Network

The proposal of Graph Convolutional Networks (GCN) gives the neural network a better performance to understand graph structure data. Therefore, GCN has a natural advantage in the action recognition task. The traditional adjacency matrix was generated depending on the physical connection of human joints. It limits the ability of GCN to understand skeleton data. In this paper, we propose a new adaptive method to generate the graph adjacency matrix and a new GCN unit to process the adjacency matrix. A series of experiments have been implemented to verify the practicability and potential of the proposed method.