Separable large-kernel graph convolution with hierarchical aggregation transformer for skeleton-based action recognition

Skeleton-based action recognition has gained significant attention due to its robustness and efficiency in understanding human motions. Recent methods leveraging Graph Convolutional Networks (GCNs) and Transformers have achieved remarkable performance. However, GCNs often struggle to capture long-range spatial dependencies, while Transformers may overlook fine-grained temporal correlations. To address these limitations, we propose a novel framework that combines a separable large-kernel graph convolutional network with a hierarchical aggregation Transformer. The separable large-kernel convolution operator expands the receptive field and improves channel adaptability in both GCN and Temporal Convolutional Network (TCN) modules, effectively modeling long-term spatial configurations and temporal dynamics. Meanwhile, the hierarchical aggregation Transformer enhances the ability to capture both local details and global patterns of skeletal sequences across multiple feature scales. Our method has been proven to surpass current leading approaches through extensive experiments on large-scale datasets, showcasing its effectiveness and generalization for improving skeleton-based action recognition