Intra-Inter Region Adaptive Graph Convolutional Networks for skeleton-based action recognition

To effectively capture the spatio-temporal dependencies of the skeletal data, graph convolution has been widely applied. However, it tends to emphasize the dependence relationship between adjacent joints and does not consider long-distance dependence relationships among joints. Another problem is single-structure temporal convolution, which is difficult to extract global temporal features. To address the above issues, we propose Intra-Inter Region Adaptive Graph Convolutional Networks (IIR-AGCN), which models the long-distance relationships of skeleton sequences at temporal and spatial scales. Our contributions are three-fold: First, to enhance global topological learning capabilities of graph convolution, we propose a regional-coupled attention module, which divides joint features into multiple sub-regions, and then constructs coupled relationships between intra-inter regions by self-attention mechanism, which realizes global joint information interaction. Second, to address the issue of difficulty in extracting global temporal features, we propose a pyramid multi-scale temporal module that extracts deep multi-scale temporal features and implements adaptive cross-scale feature fusion. Third, IIR-AGCN adopts a two-stream architecture, evaluating performances on two large-scale human skeleton datasets, NTU-RGB+D 60 and NTU-RGB+D 120, respectively.