Advances in 3D reconstruction

The thesis tackles the problem of 3D reconstruction of scenes from unstructured picture datasets. State of the art is advanced on several aspects: the first contribute consists in a robust formulation of the structure and motion problem based on a hierarchical approach, as opposed to the sequential one prevalent in literature. This methodology reduces the total computational complexity by one order of magnitude, is inherently parallelizable, minimizes the error accumulation causing drift and eliminates the crucial dependency from the choice of the initial couple of views which is common to all competing approaches. A second contribute consists in the discovery of a novel slef-calibration procedure, very robust and tailored to the structure and motion task. The proposed solution is a closed-form procedure for the recovery of the plane at infinity given a rough estimate of focal parameters of at least two cameras. This method is employed for the exaustive search of internal parameters, whise space is inherently bounded from the finiteness of acquisition devices. Finally, we inevstigated how to visualize in a efficient and compelling way the obtained reconstruction results: to this effect several algorithms for the computation of stereo disparity are presented. Along with procedures for the automatic extraction of support planes, they have been employed to obtain a faithful, compact and semantically significant representation of the scene as a collection of textured planes, eventually augmented by depth information encoded in relief maps. Every result has been verified by a rigorous experimental validation, comprising both qualitative and quantitative comparisons.