Superparamagnetic clustering of data: application to computer vision

The aim of clustering is to partition data according to natural classes pre sent in it. We proposed recently a method that makes no explicit assumption about the structure of the data and under very general and natural assumpt ions solves the clustering problem by evaluating thermal properties of a di sordered (granular) magnet. The method was tested successfully on a variety of artificial and real-life problems; here we emphasize its application to analyze results obtained by a novel method of computer vision. The combina tion of these two techniques provides a powerful tool that succeeded to clu ster properly 90 images of 6 objects on the basis of their pairwise dissimi larities. These dissimilarities, which constitute a highly non-metric set o f pairwise distances between the images, form the input for clustering. A h ierarchical organization of the images that agrees with human intuition, wa s obtained without assigning to the images coordinates in some abstract spa ce. (C) 1999 Published by Elsevier Science B.V. All rights reserved.