Abstract |
The correspondence problem is one of the most important problems in computational vision. Its importance stems from the fact that structure as well as motion information can be derived from correspondences between two images. Nevertheless, phenomena such as the projective distortions, the occlusions and the photometric alterations render the problem hard to solve. Furthermore the high computational complexity of correspondences, constitute many of the existing approaches inappropriate for certain applications. Correlation is a way to solve the correspondence problem resulting in dense disparity maps. However most correlation methods show poor results near depth discontinuities, occlusions and in textureless areas. In the current work we study the behavior of correlation-based stereo in order to better understand the sources of its weaknesses and to improve its robustness. Five improvements are suggested. Two of them aim at improving the disparity image near depth discontinuities, two others try to improve the behavior near textureless regions and the final one near occlusions. The suggested improvements, combined appropriately, form the basis of two proposed algorithms. The first one explicitly deals with the problem near depth discontinuities, while the second one focuses at improving the results, especially at areas with poor intensity or color variations. The two proposed algorithms have been tested on both real and synthetic stereo pairs. The comparison with other correlation based techniques shows significant improvements in effectiveness without much computational overhead.
|