New COD10K-C benchmark exposes camouflaged object detection weaknesses under real-world noise
Motion blur knocks SINet-v2 down 18.5 Dice points, but RobustCODLite holds firm.
A new benchmark, COD10K-C, systematically tests how well camouflaged object detection models handle real-world image corruptions like blur, sensor noise, weather, and compression artifacts. Built on the existing COD10K dataset, it includes 8 corruption types at 5 severity levels—40 conditions and 81,040 evaluation pairs total. Researchers evaluated four models: the standard SINet-v2, PFNet, and ZoomNet, plus a lightweight design RobustCODLite. Results show all models drop significantly, with motion blur causing the largest degradation: SINet-v2 loses 18.5 Dice points. Brightness and fog are less disruptive.
RobustCODLite, designed with corruption augmentation, a frequency-prior branch, and uncertainty-consistency loss, retains 92.3% of its clean Dice score under corruptions—far better than SINet-v2's 87.7%, ZoomNet's 84.8%, and PFNet's 84.1%. On the hardest corruptions, it matches or beats models that perform better on clean data. The team will release the COD10K-C GitHub repository to push forward robust detection research. This work highlights a critical gap: most benchmarks use clean images, while real cameras face noisy, degraded inputs—a problem that matters for surveillance, autonomous driving, and wildlife monitoring.
- COD10K-C adds 8 corruption types (blur, noise, weather, compression) at 5 severity levels—40 conditions, 81,040 pairs.
- Motion blur causes biggest accuracy drop: SINet-v2 loses 18.5 Dice points; RobustCODLite loses only 7.7 points.
- RobustCODLite uses corruption augmentation, frequency-prior branch, and uncertainty-consistency loss to retain 92.3% clean Dice score.
Why It Matters
Real-world cameras always see corrupted images—this benchmark pushes detection models to work reliably outside the lab.