Multi-Stream Adaptive GCN with Dropgraph for Skeleton-Based Action Recognition

Graph convolution networks (GCNs) have performed excellent results in skeleton-based action recognition. Regrettably, traditional GCN-based methods only learn from the original skeleton data, that is a fixed graph topology may not be suitable for the diversity of samples and the hierarchy of the GCN models. Another common problem with GCNs is overfitting. In the work, we propose a new multi-stream adaptive GCN with dropgraph (MS-ADGCN). On the input data the graph topology of the model learns both uniformly and individually, which increases the versatility of the model. Moreover, the spatial- temporal-channel (STC) attention module helps the model pay more attention to important features. DropGraph can discard features of related vertexes to enhance the regularization effect. Additionally, we train the joints and bones information with their motion information to further improve the model performance. We conduct massive experiments on the NTU RGB+D and Kinetics-Skeleton datasets and outperform the state-of-the-art methods.