A problem of registering many data sets - whose correspondences can be determined by cross correlation, by automatically finding landmarks or by other methods - is the possible ambiguitiy of the pairwise correspondences. Quite often, the decision of which correspondences are correct and which are outliers can only be determined in the actual process of registering the datasets.
A standard approach is to use the RANSAC algorithm. However, this algorithm doesn't converge very well. The goal of this research project is to find and analyse extensions of the RANSAC algorithm or alternative methods (e.g. graph based approaches) which are suitable for the stitching problem and the landmark based registration.