Fundamental matrix and slightly overlapping views

This paper proposes a method to compute the fundamental matrix from slightly overlapping views. Slight overlaps are preferable to substantial overlaps in large multi-camera surveillance applications, because otherwise the number of cameras and thus the overall application costs would gravely increase. Instead of using point correspondences alone, we additionally use the infinite homography. The infinite homography can be computed from line segments when we have a Manhattan world scene and assume cameras with square pixels and the same alignment of the scene in both views. All assumptions are not too restrictive and occur frequently. The infinite homography and one of the epipoles determine the fundamental matrix. The paper demonstrates a simple solution to compute the epipoles from at least two point correspondences without requiring points uniformly distributed within the images. Experiments on synthetic images clearly show the advantages of the proposed method compared to the classical eight point algorithm. Images of an office scene show the idea on real world images