Adaptive Visual Regions Categorization with Sets of Points of Interest

Abstract

The Query By Visual Thesaurus (QBVT) paradigm has strongly contributed to the visual information retrieval objective when no starting example is available. The Visual Thesaurus is a representative summary of all the visual patches in the database. Its reliable construction helps the user expression a ”mental image” by composing the visual patches according to the details he has in mind. In this paper, we introduce a relational clustering algorithm (CARD) to build the Visual Thesaurus from regions finely described by variable signature dimensions. The resulting visual categories depict the variability of regions based on local color points of interest. Therefore, we extend first the notion of image matching to regions using non-traditional metrics suitable for the multi-dimensional variables. We also, introduce an appropriate relational clustering for regions categorization using the similarity matrix induced by the latter metrics. Moreover, we propose an efficient method to speed up distance computation and reduce the feature representatives based on adaptive clustering. Our approach was tested on generic images and gives perceptually relevant visual categories.

This work was supported by the European Commission under Contract FP6-001765 aceMedia, partial support of this research was provided by an NSF-INRIA international collaboration award No. OISE-0528319.