Doctor's Theses (authored and supervised):
"Accurate and Robust Stereoscopic Matching in Efficient Algorithms";
Supervisor, Reviewer: R. Sara, M. Gelautz;
Faculty of Electrical Engineering, Czech Technical University in Prague,
oral examination: 2009-06-16.
The thesis studies dense stereoscopic techniques which are usable for accurate, robust and fast matching of high-resolution images of complex 3D scenes.
The main contributions are: (1) Image sampling invariant and affine insensitive complex correlation statistic (CCS) which is based on representing the image point neighbourhood as a response to complex Gabor lters. The CCS is a complex number with a magnitude of invariant similarity and a phase of estimated maximum position between pixels. (2) Methods for refining a disparity to sub-pixel precision - as an outcome of CCS phase, and alternatively also as a single continuous optimization problem based on a simple quadratic criterion. (3) A fast matching algorithm which avoids computing correlations for the entire disparity space by growing promising correspondence hypotheses from initial (even random) seeds. The growth is coupled with a confidently stable matching algorithm by Sara, ECCV 2002, which robustly selects the matching among competing hypotheses. (4) An algorithm for verification of given correspondences by uncalibrated dense matching. It is destined for selecting correspondences before RANSAC in challenging matching problems, with low ratio of inliers, in cluttered scenes where standard descriptor-based approach fails. An efficient procedure driven by Wald's sequential decision process grows a given correspondence while collecting statics until the decision based on learned models.
Some methods presented in the thesis go beyond the scope of 3D reconstruction, and they are applicable in many problems where the correspondences between images are sought.
Created from the Publication Database of the Vienna University of Technology.