Year of Publication


Document Type





Computer Science

First Advisor

David Nister


3D reconstruction from stereo/range image is one of the most fundamental and extensively researched topics in computer vision. Stereo research has recently experienced somewhat of a new era, as a result of publically available performance testing such as the Middlebury data set, which has allowed researchers to compare their algorithms against all the state-of-the-art algorithms. This thesis investigates into the general stereo problems in both the two-view stereo and multi-view stereo scopes. In the two-view stereo scope, we formulate an algorithm for the stereo matching problem with careful handling of disparity, discontinuity and occlusion. The algorithm works with a global matching stereo model based on an energy minimization framework. The experimental results are evaluated on the Middlebury data set, showing that our algorithm is the top performer. A GPU approach of the Hierarchical BP algorithm is then proposed, which provides similar stereo quality to CPU Hierarchical BP while running at real-time speed. A fast-converging BP is also proposed to solve the slow convergence problem of general BP algorithms. Besides two-view stereo, ecient multi-view stereo for large scale urban reconstruction is carefully studied in this thesis. A novel approach for computing depth maps given urban imagery where often large parts of surfaces are weakly textured is presented. Finally, a new post-processing step to enhance the range images in both the both the spatial resolution and depth precision is proposed.