Jun Chu, Xiaoping P. Liu, Chunlin Jiao, Jun Miao, and Lu Wang
Multi-view reconstruction, scale invariant feature, Cayley trans-forms, relaxation iterative optimization
The paper addresses the problem of large-scale multi-view registration of depth map calculated via stereo vision. In this paper we propose a fully automatic registration technique using 2D-image features for finding the corresponding points on the 3D views. We first describe the coarse pair-wise registration on neighbouring views that can estimate the rotation and translation matrix independently and retain the orthogonality of the rotation matrix. We then describe a global relaxation iteration algorithm for optimizing coarse registration to circumvent global program of accumulated errors and incrementally registries views against a growing global union of view points. The optimization procedure is based upon the condition that the product of all the transform matrices in turn is an identical transform, so its application does not limit to the panoramic annular 3D scene registration. The relaxation factor is the key to our algorithm. When we reconstruct a panoramic annular 3D scene through the optimal transform matrices, the coordinate errors of registered local models are distributed to every jointed region averagely. So the recovered panoramic 3D scene is relatively ideal. And the relaxation iteration algorithm does not require point correspondences between views, and can be used to integrate any method of pair-wise registration or robot odometry. Our experiments show that the global relaxation iteration algorithm is feasible.
Important Links:
Go Back
