On what it means to see, and what we can do about it

From Santa Fe Institute Events Wiki

Revision as of 16:24, 5 September 2007 by Mm (talk | contribs)

On what it means to see, and what we can do about it*

Shimon Edelman

Cornell University

A decisive resolution of the related problems of object recognition and categorization is at present impeded not by a shortage of computational ideas for processing the array of measurements with which vision begins, but rather by certain key tacit assumptions behind the very formulation of these problems. Recognition, which is typically taken to mean determining whether the input is a manifestation of a known entity, indiscriminately individuates and reifies such entities as visual "objects" without much concern about the nature of objecthood. More generally, categorization extends this fallacy to concepts of various degrees of abstraction, which are assumed to possess well-defined visual manifestations. Computer vision systems often further exacerbate their predicament by attempting to deal with concepts, which in human cognition form a situated, interdependent, heterarchical network, out of context, and by forcing conceptual interpretations on inputs that may be best left uninterpreted in the traditional sense. Not surprisingly, existing methods for recognition and categorization work well only for very few kinds of "objects" when confronted with realistic visual scenes, both in natural and in constructed environments. To make progress in designing robust and versatile artificial visual systems, we must start at the beginning, by considering closely the entire range of tasks, over and above recognition and categorization, that natural vision has evolved to solve. In other words, we must sooner rather than later face up to the question of what it means for us to see. Arguably, the integrated behavioral, neurobiological, and computational understanding of primate vision developed in the past two decades goes a long way towards answering this question.

  • Abstract of a chapter invited for the volume "Object Categorization:
 Computer and Human Vision Perspectives", S. Dickinson, A. Leonardis,
 B. Schiele, and M. J. Tarr, eds. (Cambridge University Press, to

Back to Program