A graph-based clustering method and its applications

Abstract

In this paper we present a graph-based clustering method particularly suited for dealing with data that do not come from a Gaussian or a spherical distribution. It can be used for detecting clusters of any size and shape, without the need of specifying neither the actual number of clusters nor other parameters. The method has been tested on data coming from two different computer vision applications. A comparison with other three state-of-the-art algorithms was also provided, demonstrating the effectiveness of the proposed approach. © 2007 Springer-Verlag Berlin Heidelberg.