Local homography estimation using keypoint descriptors

This chapter presents a novel learning-based approach to estimate local homography of points belong to a given surface and shows that it is more accurate than specific affine region detection methods. While others works attempt this by using iterative algorithms developed for template matching, our method introduces a direct estimation of the transformation. It performs the following steps. First, a training set of features captures geometry and appearance information about keypoints taken from multiple views of the surface. Then incoming keypoints are matched against the training set in order to retrieve a cluster of features representing their identity. Finally the retrieved clusters are used to estimate the local homography of the regions around keypoints. Thanks to the high accuracy, outliers and bad estimates are filtered out by multiscale Summed Square Difference (SSD) test.