Robust Stereo using Correlation Scale Space (bibtex)

by Christian Menard

Abstract:

The stereo analysis method is similar to the human visual system. Due to the way our eyes are positioned and controlled, our brains usually receive similar images of a scene taken from nearby points of the same horizontal level. Therefore the relative position of the images of an object will differ in the two eyes. Our brains are capable of measuring this difference and thus estimating the depth. Stereo analysis tries to imitate this principle. This work contains two complementary and original contributions, one combines stereo techniques with robust statistics and the other solves the correspondence problem in a multi-scale approach using correlation scale-space. Most standard algorithms make unrealistic assumptions about noise distributions, which leads to erroneous results that cannot be corrected in subsequent processing stages. In this work the standard area-based correlation approach is modified so that it can tolerate a significant number of outliers. The approach exhibits a robust behavior not only in the presence of mismatches but also in the case of depth discontinuities. Another central problem in stereo matching using correlation techniques lies in selecting the size of the search window. Small windows contain only a small number of data points, and thus are very sensitive to noise and therefore result in false matches. Whereas large search windows contain data from two or more different objects or surfaces, thus the estimated disparity is not accurate due to different projective distortions in the left and the right image. In this work a new method is proposed providing a continuous scale for the matching process, so that for each region in the stereo pair depending on the local information an optimal scale can be estimated. Results are given on synthetic images for the robust correlation technique. The adaptive matching method using correlation scale-space is tested on both synthetic and real images.

Reference:

Robust Stereo using Correlation Scale Space (Christian Menard), Technical report, PRIP, TU Wien, 1996.

Bibtex Entry:

@TechReport{PTR-Menard96a, author = "Christian Menard", institution = "PRIP, TU Wien", number = "PRIP-TR-045", title = "Robust {S}tereo using {C}orrelation {S}cale {S}pace", year = "1996", url = "ftp://ftp.prip.tuwien.ac.at/pub/publications/trs/tr45.ps.gz", abstract = "The stereo analysis method is similar to the human visual system. Due to the way our eyes are positioned and controlled, our brains usually receive similar images of a scene taken from nearby points of the same horizontal level. Therefore the relative position of the images of an object will differ in the two eyes. Our brains are capable of measuring this difference and thus estimating the depth. Stereo analysis tries to imitate this principle. This work contains two complementary and original contributions, one combines stereo techniques with robust statistics and the other solves the correspondence problem in a multi-scale approach using correlation scale-space. Most standard algorithms make unrealistic assumptions about noise distributions, which leads to erroneous results that cannot be corrected in subsequent processing stages. In this work the standard area-based correlation approach is modified so that it can tolerate a significant number of outliers. The approach exhibits a robust behavior not only in the presence of mismatches but also in the case of depth discontinuities. Another central problem in stereo matching using correlation techniques lies in selecting the size of the search window. Small windows contain only a small number of data points, and thus are very sensitive to noise and therefore result in false matches. Whereas large search windows contain data from two or more different objects or surfaces, thus the estimated disparity is not accurate due to different projective distortions in the left and the right image. In this work a new method is proposed providing a continuous scale for the matching process, so that for each region in the stereo pair depending on the local information an optimal scale can be estimated. Results are given on synthetic images for the robust correlation technique. The adaptive matching method using correlation scale-space is tested on both synthetic and real images.", }

Powered by bibtexbrowser