New similarity measures of localities for a two-layer matching scheme and estimation of fundamental matrices

Abstract

Estimation of fundamental matrices is important in 3D computer vision. It is well known that the estimation of fundamental matrices is sensitive to outliers—even a few of imprecise point correspondences may result in an estimated fundamental matrix inconsistent with the geometry setup of input images. In terms of interest points with localities, we have proposed a two-layer matching scheme that is a generalization of conventional normalized cross correlation (NCC), aiming to improve the precision of point correspondence. A locality of interest points means a set of interest points that are contiguous to each other in terms of 8-connectivity. The first layer of the matching scheme establishes locality correspondence, and the second layer refines point correspondence within matching localities. In this paper, we analyze a limitation of the similarity measure of localities proposed in our previous work and then we propose two new similarity measures to address the limitation. We test the two-layer matching scheme on images in Middlebury stereo datasets, with known fundamental matrices as the ground truth. From the experimental results, we observe the improvement of new measures of locality similarity, and we also observe that the estimated fundamental matrices are consistently close to the ground truth.