Real-world examples for constraining epipoles using point-line correspondences: self-contained benchmarks for epipole computation

Epipoles are important geometric entities of epipolar geometry, which are defined as the image of the camera center of one view in the other view. Many different algorithms in computer vision rely on the computation of epipoles, thereby giving rise to the need for efficient methods for computation of epipoles. In response to this need, different methods for either constraining or locating epipoles have been devised.

This paper exploits a special kind of correspondences across two views in order to propose a novel approach for both constraining and computation of epipoles depending on the number of correspondences. The most important property of this kind of correspondence, called point-line correspondence, is that it can be used directly for constraining the position of epipoles as soon as it is identified, eliminating the need for further processing. Unfortunately, this significant advantage has been obtained at a high cost: rarity of the desired correspondences. In order to dim this latter point, however, the paper also considers an important application of the proposed method, which is the generation of self-contained benchmarks for epipole computation. By self-contained it is meant that no information in addition to the images themselves, should be provided.