Reproduce “MIORe & VAR-MIORe: Benchmarks to Push the Boundaries of Restoration” 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

George Ciubotariu, Zhuyun Zhou, Zongwei Wu, Radu Timofte

Researchers from the University of Würzburg introduce MIORe and VAR-MIORe, two multi-task datasets, to address limitations in motion restoration benchmarks by providing high-resolution footage with diverse motion, combined blur types, and challenging environmental conditions. Evaluations on these datasets reveal that state-of-the-art algorithms exhibit varying performance, particularly struggling with moderate motion in video frame interpolation and certain blur levels in deblurring.

Key insights

  • The new datasets effectively expose limitations in existing restoration algorithms, with state-of-the-art deblurring models exhibiting varying robustness across different motion intensity splits and blur magnitudes.
  • Video Frame Interpolation methods surprisingly performed better under extreme motion conditions than moderate ones, suggesting a gap in handling complex, nuanced motion patterns.
  • Frequency-based deblurring methods, while potentially beneficial for fast motion, may introduce artifacts and struggle to generalize to slow-motion scenes, pointing to issues with their underlying priors.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.