| Stereo vision is a fundamental task in computer vision. It addresses the problem of reconstructing the 3D scene from two (or more) 2D images. The central problem in stereo is determining correspondence, or the matching of locations in one image to locations in the other on the basis of local properties such as pixel intensity, or other features. The key to success is resolving matching ambiguities. Although many matching constraints exist in computer vision, many of them are either heuristic or incorrect when applied to scenes with rich, natural structures. A rigorous study of matching constraints is therefore needed, and in this thesis the problem is approached geometrically. In particular, a frame based differential geometric analysis for the stereo reconstruction of curves and surfaces is analyzed and developed into a set of differential geometric constraints for both curve matching and surface dense stereo. The derived differential geometric consistency in effect exploits contextual information geometrically, and examples show it has resulted in much improved feature-based and surface dense stereo algorithms for complex natural scenes. |
