Reproduce “Assessing Image Inpainting via Re-Inpainting Self-Consistency Evaluation” 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
Tianyi Chen, Jianfu Zhang, Yan Hong, Yiyi Zhang, Liqing Zhang
Researchers from Shanghai Jiao Tong University and Cornell University developed a self-supervised evaluation framework for image inpainting using a re-inpainting self-consistency approach. This method overcomes biases of traditional metrics and operates without a ground truth, demonstrating strong alignment with human judgment.
Key insights
- Using small, scattered patch masks for re-inpainting significantly reduces variance in the results, making the consistency assessment more stable and reliable compared to large, normal masks.
- The 'First-Second' objective, comparing the initial inpainting result with its re-inpainted versions, exhibits the lowest variance and highest stability, effectively enabling reference-free evaluation.
- LPIPS is a robust sub-metric for this framework, consistently ranking inpainting methods regardless of the second inpainting network used, and showing superior alignment with human perception over pixel-based metrics.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.