SAM2 on a Diet: Unlocking Massive Potential With Minimal Data for Semi-Supervised Video Camouflaged Object Detection

Video camouflaged object detection (VCOD) poses a formidable challenge, requiring the segmentation of objects that seamlessly blend into their dynamic surroundings. While Vision Foundation Models (VFMs) have improved VCOD performance, their inherent generalization remains under-harnessed. Through an in-depth analysis, we discover that the primary bottleneck lies in certain "maladaptive parameters" that fail to adapt to camouflaged features under limited supervision. To address this, we propose DietSAM2, a data-efficient parameter fine-tuning framework designed to unlock the potential of SAM2 with minimal supervision. At its core, DietSAM2 integrates a novel Reverse SAM2 Parameter Configuration Module, which strategically reorients and amplifies the maladaptive parameters to bolster semantic understanding in a train-free manner. While effective globally, this optimization inadvertently compromises fine-grained local details. To mitigate this trade-off, we introduce a Pixel- wise Error Discrepancy Calibration Module to explicitly rectify under- and over-segmentation, preventing local ambiguities from distorting the representational space. Remarkably, with only a single-frame prompt and 80% fewer annotations, DietSAM2 achieves performance on par with fully supervised counterparts. Unlike traditional semi-supervised learning that typically relies on complex training pipelines or massive unlabeled data, this success underscores the tremendous potential of parameter-efficient adaptation in data-scarce and semi-supervised regimes. Moving forward, to formally establish this promising line of research, we pioneer the semi-supervised VCOD task and introduce the first dedicated evaluation protocol and benchmark, enabling standardized evaluation under low-cost settings.