ForegroundNet: Domain Adaptive Transformer for Camouflaged Object Detection

Camouflaged object detection (COD) from optical images intends to detect the “perfectly” concealed objects from their complicated backgrounds, which has promising prospects for facilitating widespread valuable tasks. However, COD currently still faces the challenges of avoiding confusion detection, false detection, and correctly identifying local structured details from complex surroundings. In this paper, we strive to embrace challenges toward more accurate COD. To this end, we develop a novel paradigm called ForegroundNet for the COD tasks. Intuitively, we propose a domain adaptation module (DAM) to transfer the domain invariant object information from the generic object detection (GOD) domain to the COD domain, to help ForegroundNet has a better object interpretation ability, resulting in better object location performance. Furthermore, we propose a ConvTransformer mechanism-aided decoding module (CMADM) to enhance the decoding of the local structured details in complicated surroundings. Extensive experiments, in this study, implemented on three widely used benchmarks, i.e ., CHAMELEON, CAMO, and COD10K, clearly demonstrate that the proposed ForegroundNet not only achieves substantial improvements in reducing false and missed detection but also attains advantages in local details fine detection, as compared with numerous representative state-of-the-art COD algorithms.