Reproduce “Experts-Guided Unbalanced Optimal Transport for ISP Learning from Unpaired and/or Paired Data” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Georgy Perevozchikov, Nancy Mehta, Egor Ershov, Radu Timofte
This work introduces Experts-Guided Unbalanced Optimal Transport (EGUOT), a flexible framework for raw-to-sRGB image translation that enables training high-performance Image Signal Processors using either unpaired or paired data. The approach achieves state-of-the-art performance for unsupervised methods, matching or exceeding the quality of models trained with extensive paired datasets, while demonstrating robustness to dataset outliers.
Key insights
- Unbalanced Optimal Transport (UOT) provides superior robustness to dataset outliers (e.g., blurry, over/under-exposed images) compared to traditional GANs or classic OT, preventing them from disproportionately affecting the learned transport map.
- Integrating a committee of specialized 'expert' discriminators is crucial for guiding the optimal transport mapping to achieve the high-fidelity perceptual quality required for raw-to-sRGB translation, addressing specific failure modes in color, structure, and frequency domains.
- The synergistic combination of UOT's outlier robustness and expert-guided adversarial loss enables effective unsupervised training of high-performance ISPs, thereby alleviating the significant bottleneck of paired data acquisition.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.