Three dimensional reconstruction from low textured stereo pairs

For textureless or low textured stereo image pairs, there are structural ambiguities in edge interpretations that can not be evaluated by local correspondence algorithms. Either the epipolar constraint or window-based matching, or a combination of both, will not offer sufficient insight for creating a unique reconstruction. In this paper, we propose a matching point-based hypothesis scheme. To address the unique global solution, a 3D reconstruction is created based on acceptance criteria, with classification of a sparse set of correspondences. Experiments with both synthesized images and complex natural scenes are used to demonstrate the usefulness of this method.