Abstract
Stereo correspondence algorithms have been based mainly on disparity in retinotopic position. The physiological evidence that nearly all disparity-selective neurons in V1 are orientation-selective has been underused e.g. to filter non-aligned orientations from the correspondence process. For example, matches to vertical edges in the left image are sought from vertical edges in the right image. This limited use of orientation ignores the information contained in orientation disparity, which signals deviation from the plane of fixation, and greatly limits the class of scenes for which it can work. We have developed a stereo correspondence algorithm that simultaneously computes both position and orientation disparity. It is based on the projective and the differential geometry of space curves, from which we derive a system of interactions between local receptive-field measurements in the left and right images. Thus the model generalizes to stereo the co-aligned and co-circular interactions whose functionality is attributed to the long-range horizontal interactions between orientation columns. But an important difference emerges between stereo and curvilinear grouping: experiments with the model predict an increased need for resolution in orientation for stereo. This suggests why the stereo computation is elaborated into V2 from V1, and why long-range interactions in V1 avoid the opposite ocular dominance column. Without this requirement, correspondence could be computed directly and efficiently in V1. Physiological data from V2 confirm these predictions.