Reproduce “Uncertainty Estimation for Super-Resolution using ESRGAN” 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
Maniraj Sai Adapa, Marco Zullich, Matias Valdenegro-Toro
Researchers from the University of Groningen integrated Monte Carlo Dropout and Deep Ensembles into SRGAN and ESRGAN to provide per-pixel uncertainty estimates for super-resolution outputs. They demonstrated that the generated uncertainty maps, particularly from Deep Ensembles, reliably correlate with reconstruction errors, offering users a measure of prediction trustworthiness.
Key insights
- The per-pixel standard deviation derived from uncertainty estimation methods serves as a reliable proxy for the test-time reconstruction error (Mean Absolute Error).
- Uncertainty maps effectively highlight regions in super-resolved images that contain visible errors or are inherently difficult for SR models to reconstruct due to insufficient information.
- Deep Ensembles generally maintained or slightly improved the primary SR performance (PSNR) while providing robust uncertainty estimates, whereas MCD sometimes led to a slight degradation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.