Epipolar geometry estimation for wide-baseline omnidirectional street view images

This paper presents a new robust method of epipolar-geometry estimation for omnidirectional images in wide-baseline setting, e.g. with Google street View images. The main idea is to learn new statistical geometric constraints that are derived from the feature descriptors into the model verification process of RANSAC. We show that these constraints provide more reliable matches, which can be used to retrieve correct epipolar geometry in very difficult situations. Robustness of epipolar-geometry estimation is quantitatively evaluated for omnidirectional image pairs with variable baseline. The performance of the proposed method is demonstrated using the complete pipeline of structure-from-motion with real dataset of Google Street View images.