Multimodal Fusion with Cross-Modal Attention for Action Recognition in Still Images

We propose a cross-modal attention module to combine information from different cues and different modalities, to achieve action recognition in still images. Feature maps are extracted from the entire image, the detected human bounding box, and the detected human skeleton, respectively. Inspired by the transformer structure, we design the processing between the query vector from one cue/modality, and the key vector from another cue/modality. Feature maps from different cues/modalities are cross-referred so that better representations can be obtained to yield better performance. We show that the proposed framework outperforms the state-of-the-art systems without the requirement of an extra training dataset. We also conduct ablation studies to investigate how different settings impact the final results.