Toward Bridging the Simulated-to-Real Gap: Benchmarking Super-Resolution on Real Data

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Quantitative super-resolution (SR) studies often rely on simulated data, potentially overestimating algorithm performance on real images.
• Bridging the gap between simulated and real-world SR evaluation is crucial for practical applications.

Purpose of the Study:

• Introduce the Super-Resolution Erlangen (SupER) database, the first comprehensive laboratory SR dataset with pixel-wise ground truth from all-real acquisitions.
• Provide a benchmark for evaluating SR algorithms on realistic data, addressing limitations of current simulated benchmarks.

Main Methods:

• The SupER database comprises over 80,000 images across 14 scenes, incorporating variations in sensor noise, sampling, motion, lighting, and video compression.
• 19 single-image and multi-frame SR algorithms were benchmarked using quantitative analysis on ground truth data and qualitative assessment via a large-scale observer study.

Main Results:

• The SupER database significantly surpasses existing benchmarks in both quality and quantity.
• Results indicate that SR algorithms performing best on simulated data may not achieve top performance on real-world data.
• Statistical analysis revealed agreements and discrepancies between quantitative and qualitative evaluations.

Conclusions:

• The SupER database offers a robust resource for evaluating super-resolution algorithms under realistic conditions.
• Findings suggest a need for algorithm development that accounts for real-world imaging artifacts and complexities.
• The publicly available SupER dataset will facilitate future research and development in super-resolution technology.