Text-guided camouflaged object detection

Journal article

Camouflaged object detection (COD) aims to identify and segment objects hidden in the background due to their high similarity in colour or texture. Recent efforts have investigated the utilization of foundation models to provide extra supervision for solving this task, such as object masks predicted by Segment Anything Models and labelled camouflaged object images generated by Stable Diffusion Models. In this work, instead of visual supervision, we endeavour to utilize Multimodal Large Language Models (MLLM) to provide textual information for COD. Specifically, we propose the Text-Guided Camouflaged Object Detection (TG-COD) framework, which consists of two main stages: extracting pseudo-textual description from MLLM and integrating textual information in the COD process. Our framework leverages the knowledge of MLLM and guides the detection process with textual information. Comprehensive experiments demonstrate that our method achieves state-of-the-art (SOTA) performance and also exhibits strong generalization capability in a few-shot setting.

Authors: Chen, Z; Xue, Y; Li, Z; Torr, P; Gu, J

• Publication status: Published
• Peer review status: Peer reviewed
• Publisher: Elsevier
• Journal: Pattern Recognition
• Volume: 170
• Article number: 112058
• Publication date: 2025-07-07
• Acceptance date: 2025-06-23
• DOI: 10.1016/j.patcog.2025.112058
• EISSN: 1873-5142
• ISSN: 0031-3203
• Language: English
• Keywords: camouflaged object detection; in-context learning; multimodel fusion