024d7f84fff11dd7e8d9c510137a2381-Reviews.html

This paper examines third-order co-occurrence statistics of edges in natural images. The main finding is that triples of edges co-occur in ways consistent with smooth flows, elongated along the tangent direction. Better statistical models of the geometry of natural image structure are crucial to understanding biological vision as well as making progress in computer vision, and the authors are correct in their claim that to date most models have focused on 2nd-order relationships, particularly the joint distribution of pairs of edges. Thus the investigation of higher-order statistics is of great interest. Overall the methods seem sound.