Can Computer Vision Problems Benefit from Structured Hierarchical Classification?

Thomas Hoyoux, Antonio Jose Rodríguez-Sánchez, Justus H. Piater, Sándor Szedmák

Abstract: While most current research in the classification domain still focuses on standard “flat” classification, there is an increasing interest in a particular type of structured classification called hierarchical classification. Incorporating knowledge about class hierarchy should be beneficial to computer vision systems as suggested by the fact that humans seem to organize objects into hierarchical structures based on visual geometrical similarities. In this paper, we analyze whether hierarchical classification provides better performance than flat classification by comparing three structured classification methods – Structured K-Nearest Neighbors, Structured Support Vector Machines and Maximum Margin Regression – with their flat counterparts on two very different computer vision tasks: facial expression recognition, for which we emphasize the underlying hierarchical structure, and 3D shape classification. The obtained results show no or only marginal improvement, which questions the way the data should be exploited for hierarchical classification in computer vision.