Assessing the role of spatial relations for the object recognition task

Abstract

It has been proved that spatial relations among objects and object's parts play a fundamental role in the human perception and understanding of images, thus becoming very relevant in the computational fields of object recognition and content-based image retrieval. In this work we propose a spatial descriptor to represent topological and orientation/directional relationships, which are obtained by means of combinatorial pyramids. A combination of visual and spatial features is performed to improve the object recognition task. We ran an experiment to evaluate the expressiveness of this representation and it has shown promising results. It was performed on the benchmark ETH-80 Image Set database and we compare our approach with a state-of-the-art method recently published. © 2010 Springer-Verlag.