Reproduce “Reference-based Image and Video Super-Resolution via C2-Matching” 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
Yuming Jiang, Kelvin C.K. Chan, Xintao Wang, Chen Change Loy, Ziwei Liu
Researchers developed C2-Matching, a framework for reference-based image and video super-resolution that robustly addresses transformation and resolution gaps in correspondence matching. It achieved a 0.7dB PSNR improvement on the CUFED5 dataset for image super-resolution and boosted video super-resolution performance by over 0.3dB on REDS and over 1dB on Vid4.
Key insights
- Explicitly addressing geometric and resolution mismatches in reference-based super-resolution significantly improves the fidelity of texture transfer.
- Knowledge distillation from a teacher network trained on ideal (HR-HR) matching can effectively guide a student network in more complex (LR-HR) matching scenarios.
- Dynamic and adaptive aggregation of reference features, rather than static concatenation, better accommodates potential misalignments and enhances super-resolution quality.
- Reference images can substantially augment video super-resolution, providing richer details than solely relying on inter-frame information.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.