This is the starting page of GDB-ICP algorithm.



Our goal is an automated registration algorithm capable of aligning image pairs having some combination of low overlap, substantial orientation and scale differences, large illumination differences (e.g.\ day and night), substantial scene changes, and different modalities. Our approach starts by extracting and matching keypoints. Rank-ordered matches are tested individually in succession. Each is used to generate an initial similarity transformation accurate only in a small image region. A generalized form of the Dual-Bootstrap ICP algorithm, introduced recently for aligning retinal images, is applied to refine this initial estimate. The Dual-Bootstrap iterates steps of (1) matching and refinement, (2) region growth and (3) model selection until the initial region expands to cover the overlap between images and the estimation has converged. After convergence, if the resulting transformation passes decision criteria then it is accepted as correct. Otherwise the next initial transformation is tested. Several important innovations --- including a new generic multi-scale feature extraction technique, bidirectional multi-scale matching, estimation of nonlinear models using constraints of mixed types, and novel decision criteria --- are used to make the Dual-Bootstrap approach work for a wide range of imagery. Experimental results on a suite of 18 challenging image pairs show that the algorithm effectively aligns 15 of the 18 pairs and rejects all misalignments when all other possible pairs are tried. This algorithm substantially out-performs algorithms based on keypoint matching alone.