Lightweight Attention Graph Convolutional Network for Two-Person Interaction Recognition

Currently, significant advancements have been made in the field of two-person interaction recognition. However, many contemporary methods rely on deep neural networks, resulting in longer runtime and higher computational resource requirements, which limits their practicality and adaptability in various application scenarios. To address these issues, we introduce the lightweight graph convolutional network (LGCN) and the lightweight spatio-temporal Transformer (LST Transformer), as well as propose the lightweight attention graph convolutional network (LAGCN). The LGCN incorporates sparse matrices, low-rank approximation, and weight-sharing techniques to reduce the parameter count and computational complexity. The LST Transformer uses average pooling and spatial partial self-attention to further enhance efficiency. These innovations allow our model to maintain a competitive accuracy while drastically reducing the number of parameters and computational load. We evaluate our model on the interaction subsets of the NTU-RGB+D and NTU-RGB+D 120 datasets, achieving the lowest parameter count and computational effort compared to advanced methods, while preserving high accuracy.