Band-Mixed Edge-Aware Interaction Learning for RGB-T Camouflaged Object Detection

For the first time, we propose the RGB-T Camouflaged Object Detection task, which addresses the challenge of detecting objects that exploit both RGB color/brightness and thermal signature to blend into their surroundings. Traditional COD approaches fall short in such complex, multi-modal scenarios. To support this novel task, we introduce the Thermal-Visible Camouflaged Object (TVCO1K) dataset, covering diverse illumination and representative scene conditions for model training and evaluation. Our research reveals that edge discrepancies between RGB and thermal bands significantly impede accurate detection. Conventional COD methods, which rely on edge information as a key prior, often suffer from noise caused by band-mixed edges, resulting in imprecise object localization and segmentation. To overcome these limitations, we propose the Band-Mixed Edge- Aware Interaction Learning Network (BEI-Net), a groundbreaking architecture that integrates a novel edge-aware supervision mechanism. This mechanism effectively mitigates edge inconsistencies and suppresses noise, ensuring more accurate boundary delineation. BEI-Net further innovates via a top-down multilevel framework: it employs multi-view attention to discern bandspecific differences, integrates an adaptive key feature refinement mechanism to filter deceptive cues, and adopts cross-band and cross-layer interaction to counteract camouflage strategies. Extensive experiments on the TVCO1K dataset demonstrate that BEI-Net outperforms 21 state-of-the-art methods, validating its superiority in RGB-T COD.