On the use of spatial relations between objects for image classification

Abstract

Image classification is addressed in this paper by utilizing spatial relation of detected objects in a rule-based fashion. Instances of particular object classes are detected combining bottom-up (learn-able models based on simple features) and top-down information(object models consisting of primitive geometric shapes such as lines). The rule-based system acts as a model for the spatial configuration of objects, also providing a human interpretable justification of image classification. Experimental results in the athletic domain show that despite inefficiencies in object detection, spatial relations allow for efficient discrimination between visually similar images classes.