Seeing and understanding: representing the visual world

Most research in computational vision has concentrated on efforts to recover three-dimensional (3D) metric structure from images. This may be because research communities are influenced by fashion. Our views about the architecture of an intelligent system possessing perception, that is, its components and their relationships, are influenced by the current dominant views regarding intelligence, the mind, and brain. One view about the brain, which has been and still is widely held by neurologists, assumes that there is a separation between the “two causally linked faculties of seeing and understanding, the former a passive and the latter an active process” [Zeki 1993]. As a consequence of this view, computational vision is commonly treated as a discipline whose goal is the recovery of metric descriptions of the scene that can be utilized for reasoning. After considerable effort and many theoretical results about scene recovery, computational vision is now moving into a relatively mature stage. It is now well understood that when we make assumptions about the scene, that is, impose specific models on the visible world, recovery of the model is usually possible provided the assumptions actually hoI d. For simple models of the geometry and the physical properties of the scene, we already have working systems that are finding their way into industry. But developing vision systems for particular environm ents, although useful, does not ./.