Assessing Image Inpainting via Re-Inpainting Self-Consistency Evaluation

TL;DR

This paper proposes a novel self-supervised evaluation method for image inpainting that assesses the consistency of multiple re-inpainting passes, reducing bias and aligning well with human judgment.

Contribution

It introduces a self-consistency based evaluation paradigm that moves away from traditional reference-dependent metrics, improving the assessment of inpainting quality.

Findings

The method correlates strongly with human judgment.

It reduces bias inherent in traditional evaluation metrics.

Extensive experiments validate its effectiveness across benchmarks.

Abstract

Image inpainting, the task of reconstructing missing segments in corrupted images using available data, faces challenges in ensuring consistency and fidelity, especially under information-scarce conditions. Traditional evaluation methods, heavily dependent on the existence of unmasked reference images, inherently favor certain inpainting outcomes, introducing biases. Addressing this issue, we introduce an innovative evaluation paradigm that utilizes a self-supervised metric based on multiple re-inpainting passes. This approach, diverging from conventional reliance on direct comparisons in pixel or feature space with original images, emphasizes the principle of self-consistency to enable the exploration of various viable inpainting solutions, effectively reducing biases. Our extensive experiments across numerous benchmarks validate the alignment of our evaluation method with human judgment.