Uncertainty-Aware Cross-Modality Fusion for Visible-Infrared Object Detection

Visible-infrared object detection is critical for a wide range of applications, e.g., surveillance, traffic management, and rescue. This is because visible and infrared images provide a complementary environmental perception, making it robust to severe illumination changes. Previous works explore fusion strategies to fully leverage the advantages of both modalities and overcome their limitations. In this paper, we propose an uncertainty-aware cross-modality fusion method for visible-infrared object detection. Specifically, we employ an attention-based module to integrate cross-modality features. In addition, we introduce an uncertainty-aware loss to guide the network to learn modal-specific features. Experiments on two benchmarks (i.e., DroneVehicle, and M3FD) show that the proposed method outperforms previous methods in terms of mAP.