Title: COD10K-C: Evaluating the Resilience of Camouflaged Object Detection Against Natural Image Distortions

Abstract:

While camouflaged object detection has seen significant advancements, existing standard benchmarks primarily assess model performance on pristine images. This approach lacks realism, as actual camera captures are frequently compromised by blur, sensor noise, weather-related interference, and compression artifacts. To address this gap, we introduce COD10K-C, a new benchmark for corruption robustness derived from the COD10K dataset. This benchmark encompasses eight distinct types of corruptions across five severity levels, resulting in 40 unique conditions and a total of 81,040 evaluation pairs.

We tested three widely used camouflaged object detection models—SINet-v2, PFNet, and ZoomNet—alongside a lightweight architecture named RobustCODLite. Our findings indicate that all models experience notable performance declines when exposed to corrupted images. Specifically, motion blur and Gaussian blur proved to be the most detrimental factors, with SINet-v2 suffering an 18.5-point drop in Dice score under motion blur conditions. In contrast, distortions such as fog and brightness adjustments had a comparatively minor impact.

RobustCODLite was designed with specific enhancements, including corruption augmentation, a frequency-prior branch, and an uncertainty-consistency loss. These features enabled it to maintain 92.3% of its clean Dice score even under corruption, outperforming SINet-v2 (87.7%), ZoomNet (84.8%), and PFNet (84.1%). Notably, on the most severe corruption levels, RobustCODLite either matched or surpassed the performance of models that achieved higher scores on clean data. To facilitate ongoing research into robust camouflaged object detection, we will publicly release the COD10K-C GitHub repository.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC