Multi-polyhedron reconstruction in a three-view system using relaxation

Abstract In this paper, two stereo matching algorithms based on relaxation are proposed for finding the correspondences among multi-polyhedron objects in a three-view system. In the first algorithm, the feature points on the second image near the intersection of the two epipolar lines derived from the first and the third images are considered as possible correspondences. In the second algorithm, the features near either epipolar line, instead of the features near the intersection of the two crossing epipolar lines as in the first algorithm, are considered as possible correspondences. The initial probability for every triplet is assigned according to the distance from the intersection of the two crossing epipolar lines (or one of the two epipolar lines) to the feature location. Connection relationship between feature points is then used in the relaxation model to resolve the ambiguious correspondence triplets. Experimental results show that the first matching algorithm is faster and requires smaller memory, while the second matching algorithm is slower but more reliable. For acute angles smaller than 60°, the first algorithm becomes unreliable and one has to appeal to the second algorithm. By properly choosing one of the two algorithms proposed, 3D multiple polyhedra in a complex scence can be successfully reconstructed with arbitrary camera configurations in a three-view system.