Dataset Card for UVE-Subjective-Benchmark

🌊 Dataset Summary

The evaluation of Underwater Video Enhancement (UVE) remains a significant challenge due to the complex visual degradations inherent to aquatic environments and the temporal instability frequently introduced by frame-wise processing. Existing objective metrics (e.g., UIQM, UCIQE) often correlate poorly with human subjective judgement, particularly regarding dynamic artifacts like flickering and color shifts.

This dataset provides a validated, large-scale subjective benchmark for UVE. It captures the genuine spatial-temporal trade-offs made by the Human Visual System (HVS), offering a mathematically sound perceptual ground truth for evaluating enhancement algorithms and training future Video Quality Assessment (VQA) models.

📊 Dataset Structure

The repository is structured to provide both the visual stimuli and the exhaustive numerical scoring matrices required for comprehensive analysis.

1. Video Data

• Standardized Sequences: A complete set of standardized underwater video sequences.
• Model Outputs: Includes the original degraded reference videos and the comprehensive enhanced outputs generated by 15 representative deep learning models.
• Taxonomy Covered: Evaluated models span multiple architectural families, including physics-based priors, CNNs, GANs, Vision Transformers, and a dedicated video architecture.

2. Evaluation Matrices

• Objective Metrics: Frame-by-frame UIQM and UCIQE objective scores for all enhanced sequences.
• Aggregated Data: Video-level and algorithm-level average scores.
• Subjective Ground Truth: Reconstructed subjective Bradley-Terry (BT) scores and all associated statistical metrics derived from human preference data.

🔬 Methodology

To neutralize observer scale drift and effectively isolate dynamic artifacts, the human preference data was collected utilizing a strict, forced-choice Pair Comparison (PC) methodology. A bespoke, zero-latency web-based evaluation platform was engineered to facilitate synchronized, high-resolution video playback and collect robust PC data.

The discrete, binary human choices were then mathematically transformed into a continuous perceptual quality ranking using the Bradley-Terry (BT) probabilistic model and Maximum Likelihood Estimation (MLE).

📝 Citation

If you use this dataset in your research, please cite the underlying work:

@article{Wang2026UVE,
  title={Quality Assessment on Enhanced Underwater Video},
  author={Wang, Tianshuo},
  year={2026},
  institution={Central South University & University of Dundee}
}