Geometric Constraints In Stereo Vision

The correspondence problem, matching the same feature in two views, is a central problem of stereo vision. We examine geometric constraints on stereo correspondence and describe progress toward formulation from first principles of an evaluation function for selection of the best among alternative correspondences. Within a surface interpretation, conditions on correspondence of edges and surface intervals are shown. These conditions are useful with wide angle stereo and provide particularly tight constraints for narrow angle stereo. We invoke the general assumption that edge and surface orientations are not related to observer position. We are combining these general constraints with ongoing work in scene modeling of known regularities which include distributions with respect to the gravity vector (horizontal and vertical edges and surfaces), parallelism and alignment with local coordinate systems, and orthogonal corners. We will use them to calculate a likelihood measure for correspondences.