Towards Language-Guided Unified Visual Prompt Learning for Camouflaged Object Segmentation

Camouflaged object segmentation (COS) aims to segment objects that seamlessly blend into the background, making it a crucial task in visual perception. Most existing COS methods formulate this problem in a purely visual manner, and thus struggle to capture the semantic cues and contextual knowledge required for understanding camouflage. Recently, language-guided segmentation methods have offered a promising solution. To advance language-guided COS, we study Language-Guided Unified Visual Prompt Learning for COS, where camouflaged objects are segmented under text-prompt guidance with diverse visual-prompt conditions, including mask, point, scribble, and null-prompt settings. To support this setting, we re-annotate existing COS benchmarks by collecting 12,487 text descriptions and constructing corresponding visual prompts, including full-mask, point, and scribble annotations, together with a no-prompt protocol. Building on this, we propose a unified COS framework based on a multimodal large language model (MLLM), which incorporates different visual prompts under text-prompt guidance. Extensive experiments show that our framework achieves competitive performance across different visual prompt settings.