Fusion Attention for Action Recognition: Integrating Sparse-Dense and Global Attention for Video Action Recognition

Conventional approaches to video action recognition perform global attention over the entire video patches, which may be ineffective due to the temporal redundancy of video frames. Recent works on masked video modeling adopt a high-ratio tube masking and reconstruction strategy as a pre-training method to mitigate the problem of focusing on spatial features well but not on temporal features. Inspired by this pre-training method, we propose Fusion Attention for Action Recognition (FAR), which fuses the sparse-dense attention patterns specialized for temporal features with global attention during fine-tuning. FAR has three main components: head-split sparse-dense attention (HSDA), token-group interaction, and group-averaged classifier. First, HSDA splits the head of multi-head self-attention to fuse global and sparse-dense attention. The sparse-dense attention is divided into groups of tube-shaped patches to focus on temporal features. Second, token-group interaction is used to improve information exchange between divided patch groups. Finally, the group-averaged classifier uses spatio-temporal features from different patch groups to improve performance. The proposed method uses the weight parameters that are pre-trained with VideoMAE and MVD, and achieves higher performance (+0.1-0.4%) with less computation than models fine-tuned with global attention on Something-Something V2 and Kinetics-400. Moreover, qualitative comparisons show that FAR captures temporal features quite well in highly redundant video frames. The FAR approach demonstrates improved action recognition with efficient computation, and exploring its adaptability across different pre-training methods presents an interesting direction for future research.