A methodological approach to visual reasoning

The lack of a structured and theoretical approach to reasoning with visual information is a major obstacle for designing and implementing vision systems. The authors provide a four-step methodological approach to this problem: (1) visual reasoning problems are classified into several categories by their characteristics; (2) a particular class of problems, namely two-dimensional object recognition, is examined, and a system model for describing it is formally extracted; (3) several probability-related inference mechanisms are examined, and the one that is most appropriate for reasoning with the above model is chosen; and (4) a hybrid inference mechanism, the outcome of the above comparative study, is applied to a visual reasoning problem. They show how this approach can be used to overcome the difficulties in designing visual reasoning systems. The implications of this work are discussed