Reproduce “Video Deblurring by Fitting to Test 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
Xuanchi Ren, Zian Qian, Qifeng Chen
The HKUST research introduces a "fitting-to-test-data" pipeline for video deblurring, training a network directly on the blurry input video to overcome the domain gap inherent in pre-trained models. This self-supervised approach, accelerated by MAML, significantly improves deblurring quality with user preference rates between 66.0% and 94.2% while reducing processing time from hours to approximately five minutes per video.
Key insights
- Training a model directly on a single test video can effectively address the domain gap for deblurring, leading to highly adaptive and robust performance.
- Even blurry videos contain sufficient internal information (sharper frames, scene textures) that can be leveraged for self-supervised deblurring without external datasets.
- Meta-learning can significantly reduce the computational burden of internal learning approaches, making them practical for real-world applications with minimal quality compromise.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.