Computer Vision Systems: Past, Present, and Future

Human beings are equipped with marvelous biological resources to enable interpretation of visual stimuli. Since the early sixties researchers in computer vision have been trying to teach computers to perform the same kind of tasks that humans do so well. Early systems worked exclusively in the “blocks world” domain, trying to separate out and identify each polyhedron in a scene. The use of constraint analysis was introduced and physical constraints on edges and vertices were developed for polyhedral objects and later extended to include objects having curved surfaces. Blocks world objects were essentially modeled by surface-edgevertex representations which did not easily extend to more complex objects. The emphasis switched to relational models employing three-dimensional primitives. The generalized cylinder was introduced and successfully used as the building block of models in a number of systems. At the same time, advanced control mechanisms such as pyramid structures and discrete and continuous relaxation processes were being used. We are now at a stage where full-blown vision systems employing many levels of cooperating processes are being built. In this paper we will review the important earlier systems and give detailed reports of the more recent ones. We will then predict the direction of future systems.