COMPUTING SHAPE USING A THEORY OF HUMAN STEREO VISION

This thesis is concerned with the construction of a complete specification of the shapes of the visible surfaces of a scene, given a stereoscopic pair of images of that scene as input. This problem is investigated within the framework of a computational theory. There are two parts to the problem. A theory of human stereo vision was recently developed by Marr and Poggio (1979). Part I of this thesis deals ith a computer implementation of that theory. The implementation served both as a source of feedback for the theory itself, outlining problems and illuminating previously overlooked difficulties, and as a method of testing the adequacy of the theory. By running the program on a series of images from the human stercopsis literature and comparing the results with the perceptions of stereo observers, it was shown that the program, as a reflection of the Marr-Poggio theory, adequately accounted for a wide range of stereoscopic perception. Part I1 deals with a computational theory of surface interpolation. One of the consequences of the Marr-Poggio theory, and of most early visual processes, is that explicit surface information is available only at a certain restricted set of positions in the image. In order to construct a complete specification of the surface shapes, it is necessary to account for implicit information in the formatio