On feature point matching, in the calibrated and uncalibrated contexts, between widely and narrowly separated images

Abstract

In this work, the correspondence problem for feature points between images is investigated. In this context, two important factors greatly influence the choice of a strategy: whether the camera system is calibrated or not, and how large is the separation between viewpoints. This work is divided into four parts, for the four important matching situations generated by these two factors. In the case of uncalibrated narrowly separated views, a framework for empirically evaluating matching constraints is presented. Then, various new and existing constraints are compared. In the case of calibrated narrowly separated views, a new type of feature is introduced, epipolar gradient features. These are then shown to be especially appropriate for matching in the context of quick reconstruction. The features are then matched with a new constraint based on trinocular line transfer. In the case of uncalibrated widely separated views, it is shown how the shape of feature points can be used to recover local perspective deformation between two views, and improve matching results. To this end, a new corner detector that generates the required information is also introduced. In the case of calibrated widely separated views, a more accurate estimate of local perspective deformation is obtained by incorporating the knowledge of the epipolar geometry. An application to fundamental matrix estimation is also introduced.