Fig. 2

DASF Module Architecture. During training, DASF receives RGB and IR features from modality-specific branches and generates fused features via two stages: (1) Channel Switching with AKSA adaptively exchanges cross-modal information at the channel level, producing modality-switched features \({\bar{F}}_v\) and \({\bar{F}}_r\); (2) Spatial Attention fuses the switched features through channel-wise average/max pooling and element-wise multiplication, producing the final fused output \(F_{\text {fused}}\). This fused representation provides gradient-based regularization, guiding modality-specific branches toward cross-modal invariance through backpropagation. At inference, DASF is inactive.