Computational Vision: A Probabilistic View of the Multi-Module Paradigm

Abstract

Computational Vision may be considered as the study of visual perception from an abstract viewpoint —that is, as the study of the general computational problems and solution strategies that must be solved by any organism (or robot) endowed with visual abilities. One basic strategy that has guided research in this field has been to perform psychophysical experiments that identify specific isolated computational problems that the human visual system solves; thus, it has been found that one can extract relevant information from images where all “high-level cues” (i.e., recognizable objects) have been removed (see, for example the work of Julesz (1981-a and b) on stereo and texture; the work of Land, 1971, etc.). These experiments have suggested a plausible organization for the visual system in which:

• i) Segmentation of the visual scene with respect to different attributes (such as: stereo disparity, image motion, texture, lightness, color, etc.) as well as the computation of specific geometric properties of the scene (such as relative depth and shape) occur before recognition or reasoning take place. Such purely perceptual processes have been grouped under the name of “low-level vision”.

• ii) There are individual computational modules that carry out these tasks in a more or less independent way.