NeurIPS 2020

An Unsupervised Information-Theoretic Perceptual Quality Metric

Meta Review

This paper proposes a novel perceptual image quality evaluation metric based on unsupervised method that aims optimization of a lower bound of the multivariate mutual information. The method is implemented using deep neural networks and tested two datasets, BAPPS and ImageNEt-C and was shown to achieve results comparable with supervised methods. The approach is well-motivated and clearly presented, and tackles an important problem without requiring subjective human judgements. One of the reviewers raise the question of applicability of the approach on compressed images. Others also suggest the experimentation part of the paper is somewhat preliminary.