Diffusion-Based Camouflaged Object Segmentation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Camouflaged Object Detection (COD) is challenged by the annotation bottleneck: obtaining pixel-level masks for structurally diverse, low-contrast scenes is laborious, while deterministic CNN-based segmenters often hallucinate confidence. Diffusion-based COD methods address overconfidence by sampling multiple plausible masks per image, but existing pipelines discard this epistemic uncertainty by averaging the samples into a single mask, instead of exploiting it for improving performance. This work proposes a four-stage, uncertainty-aware active learning pipeline that repurposes this discarded variance as an information signal. A pretrained diffusion COD model performs zero-shot multi-sampling on unlabeled images; pixel-wise mask variance ranks images by structural uncertainty; human annotation is directed only at the highest-uncertainty subset, cutting labeling cost substantially; and a lightweight discriminative segmenter is trained on these sparse, high-value labels for real-time deployment. The approach targets comparable segmentation accuracy to full-data baselines while using only a fraction of the annotation budget, demonstrating that architectural resourcefulness can unite the uncertainty-awareness of generative diffusion models with the real-time inference speed of discriminative networks.