Mutual Correspondences: An Hybrid Method for Image-to-geometry Registration

Abstract

Image registration is an important task in several applications of Computer Graphics and Computer Vision. Among the large number of proposed approaches, currently there is no solution which is automatic and robust enough to handle any general case. The most robust methods usually require a significant intervention by the user to specify many 2D-3D correspondences, while automatic techniques often rely on strong assumptions about the quality of 2D and 3D data. In this paper we present Mutual Correspondences, which is based on a minimization function which combines correspondences based and Mutual Information based approaches, and takes advantage of the strong points of both. Mutual Correspondences give the user the possibility to "guide" Mutual Information with only a few 2D- 3D correspondences. The proposed approach results in a wider convergence range and in higher registration accuracy, regardless of the quality of both the image and the 3D model. Mutual Correspondences were applied on some practical cases, where state-of-the-art approaches tended to fail, and they provided a mean to obtain accurate results. This led to a simple, robust and practical approach that can provide a way to register images in a few seconds.