A Study of the Region Covariance Descriptor: Impact of Feature Selection and Image Transformations

Abstract

We analyse experimentally the region covariance descriptor which has proven useful in numerous computer vision applications. The properties of the descriptor--despite its widespread deployment--are not well understood or documented. In an attempt to uncover key attributes of the descriptor, we characterise the interdependence between the choice of features and distance measures through a series of meticulously designed and performed experiments. Our results paint a rather complex picture and underscore the necessity for more extensive empirical and theoretical work. In light of our findings, there is reason to believe that the region covariance descriptor will prove useful for methods that perform image super-resolution, deblurring, and denoising based on matching and retrieval of image patches from an image dictionary.