Random sampling methods for two-view geometry estimation

TL;DR: This thesis treats efficient estimation algorithms for the epipolar geometry, the model underlying two views of the same scene or object, by investigating techniques for faster fundamental matrix estimation using RANSAC.

Abstract: This thesis treats efficient estimation algorithms for the epipolar geometry, the model underlying two views of the same scene or object. The epipolar geometry is computed from image correspondences that are found by local feature matching. These correspondences are used to calculate the fundamental matrix, which is the mathematical representation of the epipolar geometry. Since there are outliers among the correspondences, the fundamental matrix is usually calculated by the robust RANSAC (RANdom SAmple Consensus) algorithm which is very well suited for this purpose. A disadvantage of the algorithm, however, is that it shows a considerable complexity for higher outlier ratios. This hampers its application in vision algorithms dealing with many views. In this thesis we investigate techniques for faster fundamental matrix estimation using RANSAC. The first approach that is taken is the computation of inlier probabilities for the correspondences, that are used during sampling in the RANSAC algorithm to stimulate the selection of inliers. The second approach is the reduction of the required number of RANSAC samples by the selection of fewer correspondences per sample. The fundamental matrix hypotheses are then completed using the remaining correspondences.