Characterizing the performance of multiple-image point-correspondence algorithms using self-consistency

Abstract

A new approach to characterizing the performance of point-correspondence algorithms is presented. Instead of relying on any “ground truth’, it uses the self-consistency of the outputs of an algorithm independently applied to different sets of views of a static scene. It allows one to evaluate algorithms for a given class of scenes, as well as to estimate the accuracy of every element of the output of the algorithm for a given set of views. Experiments to demonstrate the usefulness of the methodology are presented.