Spatial Configuration Modeling for Efficient Visual Thesaurus Query Composition

The Visual Thesaurus (VT) is a novel query approach that overcomes the lack of starting example image and provides a panel of visual patches summarizing all the regions of the database. The user composes his query by arranging these patches according to his mental image. The retrieved images are displayed with respect to their photometric descriptors and their spatial configuration. This paper presents a novel approach to model the spatial relations between the composed visual patches. In this paper, we define the Weighted Angle Spatial Histogram (WASH) that jointly uses the (ir)regularity and angles between regions of interest to determine their relative spatial position. WASH has shown great robustness to region shape and scale in the image because it handles segmented regions as a composition of relevant and minor subregions. Moreover, using WASH descriptor allows efficient query composition from the Visual Thesaurus since retrieved images are constrained with the photometric and geometric attributes of selected visual patches. We tested our approach on generic database, and we compared it with other state-of-the-art techniques.