Deep Multimodal Fusion Detection through Spatial Mask and Channel Competition
The ADCR framework improves multimodal object detection by decoupling training and inference. It uses Spatial Mask Enhancement (SME) during training to encourage spatial feature exchange between modalities, preventing over-specialization. At inference, Learnable Channel Competition (LCC) arbitrates the best channel based on attention differences, making the system robust to modality asymmetry.