Medusa Synthesized

摘要

Medusa [Aloimonos, 1990] was introduced as a framework for designing vision for behavioral systems in the form of modules that are directly related to perceptual tasks. According to this framework, the understanding of perceptual information processing systems should be achieved through the modelling of the observer and the world in a synergistic manner, and not through the isolated modelling of the observer and the world as closed systems. The question remains: how should such synergistic modelling be realized? This paper addresses this question by providing a methodology for synthesizing vision systems and integrating perception and action. In particular, we outline an architecture for purposive vision systems and present a hierarchy of navigational competences based on computational models of increasing complexity. Purely computational considerations are not adequate for telling us what visual competences and representations are of importance in performing a set of tasks. Interaction with empirical sciences such as neurobiology, physiology, psychology, ethology, etc., on the other hand, can give us inspiration about the visual categories relevant to systems in real-world environments. Throughout this paper we describe biological findings and how they affect the choice of computational models and representations to be used for the synthesis of the hierarchy of navigational competences into a working system, i.e., for Medusa's synthesis.