Skeleton Sequence and RGB Frame Based Multi-Modality Feature Fusion Network for Action Recognition

Action recognition has been a heated topic in computer vision for its wide application systems. Previous approaches achieve improvement by fusing the modalities of skeleton sequence and RGB video. However, such methods pose dilemma between accuracy efficiency high complexity video network. To solve problem, we propose multi-modality feature fusion network to combine frame instead video, as key information contained combination is close that In this way, complementary retained while reduced large margin. better explore correspondence two modalities, two-stage framework introduced early stage, introduce attention module projects on single help focus limb movement regions. late cross-attention fuse exploiting correlation. Experiments benchmarks, NTU RGB+D SYSU, show proposed model achieves competitive performance compared with state-of-the-art reducing