A unified multi-stream diffusion framework for robust video camouflaged object detection.

Abstract

Video Camouflaged Object Detection (VCOD) aims to identify and segment objects in videos that blend seamlessly with their surroundings. This task remains highly challenging due to indistinct object boundaries and the scarcity of discriminative features caused by low foreground-background contrast. To address these challenges, we propose a unified multi-stream diffusion framework that leverages the strong representational capacity and iterative denoising mechanism of diffusion models to handle weak boundaries and blurred content in camouflaged scenes. A simple yet efficient Timestep Noise Adaptor (TNA) is further introduced to decouple timestep information from the denoising network, thereby alleviating timestep dependency and sensitivity. In addition, we design a Deformable Scale Integration Module (DSIM) to better capture camouflaged objects within structurally sparse backgrounds. DSIM integrates multi-scale information through Dual Differential Attention (DDA) and employs Deformable Top-K Sparse Attention (DTKSA) to adaptively adjust receptive fields, ensuring that the model consistently focuses on the most informative features for accurate camouflage detection. Experimental results demonstrate that our method achieves state-of-the-art performance across all six evaluation metrics on two widely-used benchmark datasets. Code will be released upon publication.