New PCD metric and TUB dataset tackle underwater vision in turbid scenes

Underwater computer vision models often struggle in murky conditions, partly because they rely on synthetic turbidity datasets that may not reflect true information loss. To bridge this gap, researchers introduce the Turbid Underwater Baseline (TUB) dataset — 1,320 images captured in real extreme turbidity, paired with over 16,000 high‑confidence ground‑truth segmentation masks. This real‑world benchmark is designed to evaluate how well vision models preserve structural details in degraded scenes.

Alongside the dataset, the team proposes PCD, a novel metric derived from phase congruency maps. Unlike standard metrics that are sensitive to contrast changes, PCD is contrast‑invariant and specifically captures the loss of structural information. In experiments, PCD showed strong correlation with the performance of instance segmentation models on both real and synthetic turbid images, while existing metrics failed. The dataset and code are publicly available, offering a new standard for validating underwater perception systems.