Epipolar Plane Images as a Tool to Seek Correspondences in a Dense Sequence

摘要

We present a method seeking correspondences in a dense rectified image sequence, considered as a set of Epipolar Plane Images (EPI). The main idea is to employ dense sequence to get more information which could guide the correspondence algorithm. The method is intensity based, no features are detected. Our spatio-temporal volume analysis approach aims at accuracy and density of the correspondences. Two cost functions are used, quantifying the belief that given correspondence candidate is correct. The first one is based on projections of one scene point to the spatio-temporal data, while the second one uses two--dimensional neighborhood (in image plane) around such projections. The assumed opaque Lambertian surface without occlusions allows us to use a simple correspondence seeking algorithm based on minimization of a global criterion using dynamic programming.