Debate-Enhanced Pseudo Labeling for Weakly-Supervised Camouflaged Object Detection

A new two-stage framework, D³ETOR, improves weakly-supervised camouflaged object detection (WSCOD) using scribble annotations. The method addresses two key limitations: unreliable pseudo masks from general-purpose models like SAM, which lack task-specific understanding, and annotation bias in scribbles that obscures global object structure. Stage one introduces adaptive entropy-based debate-enhanced pseudo labeling to generate more reliable masks. Stage two applies frequency-aware progressive debiasing to correct scribble bias. The approach aims to close the gap between weakly- and fully-supervised COD methods.