Stereo matching using M-estimators

Abstract

Stereo computation is just one of the vision problems where the presence of outliers cannot be neglected. 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 behaviour not only in the presence of mismatches but also in the case of depth discontinuities. Experimental results are given on synthetic and real images.