Spatial-Perception Fusion and Progressive Refinement Network For RGB-D Camouflaged Object Detection

Camouflaged object detection (COD) presents a significant challenge in accurately identifying objects that seamlessly blend well with their surroundings. Taking inspiration from natural vision systems, we suggest integrating depth images as an additional resource to enhance camouflage detection by providing spatial clues and separation for foreground and background. To address this issue, we innovatively propose Spatial-Perception Fusion and Progressive Refinement Network (SPPNet) to uncover hidden objects in RGB and depth modes, utilizing three components. Initially, we extract features at multiple scales from RGB images and depth maps by a dual stream Swim Transformer encoder. Then, we develop a spatial-perception fusion module that utilizes spatial attention mechanisms to combine features from multiple modalities and levels. Finally, we propose a progressive refinement module that integrates low-level features and refine edge details of predictions to precisely detect camouflaged objects. Our SPPNet outperforms other state-of-the-art COD models on three benchmark camouflaged datasets. Codes can be found at https://github.com/Mercurioy/SPPNet.