Unsupervised Learning of Part-Based Representations

Abstract

This article introduces a segmentation method to automatically extract object parts from a reduced set of images. Given a database of objects and dividing all of them using local color histograms, we obtain an object part as the conjunction of the most similar ones. The similarity measure is obtained analyzing the behaviour of a local vector with respect to the whole object database. Furthermore, the proposed technique is able to associate an energy to each object part being possible to find the most discriminant object parts. We present the non-negative matrix factorization (NMF) technique to improve the internal data representation by compacting the original local histograms (50D instead of 512D). Moreover, the NMF based projected histograms only contain a few activated components and this fact improves the clustering results with respect to the use of the original local color histograms. We present a set of experimental results validating the use of the NMF in conjunction with the clustering technique.

This work is supported by Comissionat per a Universitats i Recerca de la Generalitat de Catalunya and Ministerio de Ciencia y Tecnología grant TIC2000-0399-C02-01.