Monge-Kantorovich quantiles and ranks for image data

Abstract

This paper defines quantiles, ranks and statistical depths for image data by leveraging ideas from measure transportation. The first step is to embed a distribution of images in a tangent space, with the framework of linear optimal transport. Therein, Monge-Kantorovich quantiles are shown to provide a meaningful ordering of image data, with outward images having unusual shapes. Numerical experiments showcase the relevance of the proposed procedure, for descriptive analysis, outlier detection or statistical testing.