Beyond Restoration: Content-Degradation Factor Unfolding for Real-World Concealed Object Segmentation

beyond-restoration:-content-degradation-factor-unfolding-for-real-world-concealed-object-segmentation

Beyond Restoration: Content-Degradation Factor Unfolding for Real-World Concealed Object Segmentation

[Submitted on 18 Nov 2026]

View PDF HTML (experimental)

Abstract:Reliable visual perception under degradation requires distinguishing task-relevant content from degradation-induced variations, especially in concealed object segmentation (COS), where foreground cues are inherently weak. The central challenge is not merely improving visibility, but identifying reliable content evidence amid degradation-induced variations. RGB-space segmentation often entangles content and degradation cues, while restoration-based methods primarily focus on image recovery and task interaction. This motivates explicit factorization to distinguish content evidence from degradation effects and assess its reliability. Retinex provides a spatially aligned basis for this separation but remains illumination-centric. We therefore extend it to content-degradation formation with task-asymmetric factor roles. The Factor Unfolding Network (FUN) uses the content factor R as the sole source of direct semantic evidence for segmentation, while the degradation factor L captures degradation-related variations and generates a spatial reliability map to calibrate this evidence. A unified energy couples observation formation with reliability-weighted foreground-background separation, and its alternating updates yield a coarse-to-fine network without clean-image supervision. In this way, FUN factorizes degradation for perception rather than removing it. Extensive experiments across clean and degraded conditions and diverse segmentation tasks demonstrate robust performance and generalization. Code will be released.

Submission history

From: Rihan Zhang [view email]
[v1] Wed, 18 Nov 2026 22:17:58 UTC (9,728 KB)