2-D Vision

Abstract

Machine vision research has been spurred by the ease with which biological systems process visual inputs. Unfortunately, the task of understanding a scene from machine vision alone has proved to be difficult. The analogy of an image matrix to a human retina has only served to illuminate the powerful kinds of processing taking place in the visual cortex, processing that is poorly understood at present. The research of David Marr and others has tried to isolate those parts of human visual information processing that seem to operate independently, such as stereopsis, and to apply this knowledge to machine vision systems. While some progress has been made, the state of machine vision is still primitive. At present, most commercial machine vision systems are binary systems that use simple template matching of 2-D silhouettes. If the object is presented in a different pose or the lighting is such that a specularity or reflection upsets the silhouette algorithms, recognition becomes impossible. What these systems lack is a way of inferring and understanding the 3-D structure of the objects to be recognized. The human visual system has little trouble performing such tasks. We can understand and recognize the objects in a scene in the presence of noise and distortion and under a variety of different lighting conditions. We can even perceive 3-D from photographs and paintings which are inherently 2-D.