Research Seminar - September 17, 1998

Abstract

A fundamental problem faced by stereo vision algorithms is that of determining correspondences between two images which comprise a stereo pair. One class of matching algorithms, known as area-based, use matching metrics such as the Sum of Absolute Differences (SAD), Sum of Squared Differences (SSD) and Normalised Cross Correlation (NCC) to measure the similarity between regions of pixel values, and thus locate the optimum match. An extension to area-based matching techniques involves applying a local transform to the grey level pixel values prior to matching. Two non-parametric transforms which have been recently proposed for the stereo matching problem are the rank and census. The advantages of these transforms include their invariance to radiometric distortion, and their amenability to fast hardware implementation. This talk will give an overview of area based matching and the rank and census transforms, and will show some disparity results obtained. Some of the problems faced by matching algorithms will be discussed. The talk will also cover a number of techniques for identifying invalid matches, as well as methods of assigning a confidence level to a match. In addition, one scheme for predicting the effects of noise on matching will be discussed.