REPRESENTATION AND RECOGNITION IN VISION.: By Shimon Edelman. 1999. Cambridge (MA): The MIT Press. Price pound37.95. Pp. 335 + xxiii. ISBN 0-262-05057-9.

Abstract

As we move about the world and view objects from different distances and directions, the images projected onto the retina are correspondingly transformed—by translation, magnification, rotation and so on. How, despite rarely experiencing the same image twice, do we recognize an object as being the same? This problem has had a long and varied history involving disparate disciplines: philosophy and psychology; computer science, mathematics, and engineering; anatomy and physiology; and, to a limited extent, neurology. There is still no universally accepted solution. In broad terms, explanatory theories can be divided into two kinds: those that concentrate on the nature of the internal model or representation that the visual system forms of an object in the external world, and those that concentrate on how that representation alters with changes in the pose of the object or changes in observer viewpoint. The two approaches involve a trade-off. Thus, if the representation is assumed to alter with a change in viewpoint, then the visual system needs to apply an internal restoring, interpolating, or normalizing transformation for recognition to occur; and the more work that is done in forming the representation (the more invariant it is to viewpoint changes), the less work that needs to be done in normalizing it. In practice, factors such as the need to retain information about viewpoint changes and whether objects are to be compared simultaneously or sequentially influence the extent of the trade-off. One of the first significant theoretical analyses of visual (and auditory) representation and recognition was undertaken in the 1940s by Walter Pitts and Warren McCulloch, who showed how a biological system could calculate constant properties—invariants—of entire images under changes in viewpoint. It was not at all …