Classifying Images from Athletics Based on Spatial Relations

Spatial relations between image regions are used in this paper for image classification in a rule-based fashion. In the particular case where image regions correspond to semantically interpretable objects the rules provide the means for justifying classification in a human-familiar manner. In the work presented here instances of particular object classes are detected combining bottom-up (learnable 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. Experimental results in the athletic domain show that despite inaccuracy in object detection, spatial relations allow for efficient discrimination between visually similar images classes.