The performance of existing robust homography estimation algorithms is highly dependent on the inlier rate of feature point correspondences. In this paper, we present a novel procedure for homography estimation that is particularly well suited for inlier-poor domains. By utilizing the scale and rotation byproducts created by affine aware feature detectors such as SIFT and SURF, we obtain an initial homography estimate from a single correspondence pair. This estimate allows us to filter the correspondences to an inlier-rich subset for use with a robust estimator. Especially at low inlier rates, our novel algorithm provides dramatic performance improvements.
View on arXiv