| Abstract: |  This thesis is focused on correspondence verification which is one of the less researched stages of wide baseline stereo matching. A correspondence can be informally described as a pair of points of interest, one in the first image, the other in second image. During the process of wide baseline stereo matching, these correspondences can be identified as either correct or incorrect. Correct classification of correspondences has the potential to make the matching process much more effective. This thesis is focused on the performance of the Sequential Correspondence Verification algorithm introduced in \cite{scv} in combination with view synthesis \cite{mods} on scenes with significant viewpoint changes. The experiments conducted in this thesis show that the SCV is not suitable for scenes with extreme viewpoint changes. Moreover, it was discovered that the SCV is much more sensitive to precision of the input correspondences than it had been previously thought. Minor contribution of this work is the discovery that correspondences created with up-to-date software version are less accurate than the correspondences used in \cite{scv} which were generated by an older version of the same software.
|
|---|