Local Q-Convexity Histograms for Shape Analysis

摘要

In this paper we propose a novel local shape descriptor based on Q-convexity histograms. We investigate three different variants: (1) focusing only on the background points, (2) examining all the points and (3) omitting the zero bin. We study the properties of the variants on a shape and on a texture dataset. In an illustrative example, we compare the classification accuracy of the introduced local descriptor to its global counterpart, and also to a variant of Local Binary Patterns which is similar to our descriptor in the sense that its histogram collects frequencies of local configurations. We show that our descriptor can reach in many cases higher classification accuracy than the others.