Automatic structure recovery and visualization

We describe an automated pipeline for the reconstruction and rendering of three dimensional objects, with particular emphasys for urban environments. Our system can robustly recover 3D points and cameras from uncalibrated views, without manual assistance. The reconstructed structure is augmented by fitting geometrical primitives such as planes and cylinders to the sparse point cloud obtained. Such information is the key to obtain a higher level understanding of the scene; we use this knowledge to efficiently render the recovered environment, capturing its global appearance while preserving scalability. Several examples display our system in action. In this paper we describe the fully automated approach to object reconstruction from images that we developed over the last five years. Our integrated approach is composed by a geometric back- end and a graphical front-end.