InfoResearchPreprint
Do Underwater Image Quality Metrics Agree with Human Perception? A Benchmark with Human Choices, Simple Edits and Generative Drift
- Published
- Record updated
Summary
Researchers test whether 73 no-reference image quality metrics for underwater images agree with human preferences, using five reference sets and a new paired-comparison study. No metric consistently matches people; the best worst-case Cohen's kappa is 0.23 (Q-Align), while two raters agree with a median kappa of 0.41, and UCIQE often favors edits people reject. The authors also release TrustUIQA, a benchmark pairing human references, study choices, edited and generated images, and scores for every metric on every image.