AUTOMATIC BUILDING RECONSTRUCTION FROM AERIAL IMAGES : A GENERIC BAYESIAN FRAMEWORK

A novel system for automatic building reconstruction from multiple aerial images is presented. Compared to previous works, this approach uses a very generic modeling of buildings as polyhedral shapes with no overhang, in which external knowledge is introduced through constraints on primitives. Using planes as base primitives, the algorithm builds up an arrangement of planes from which a 3D graph of facets is deduced. In a so-called “compatibility graph” where the nodes are the initial facets of the 3D graph and edges between two nodes state that both facets belong to at least one common hypothesis of building, it is shown that maximal cliques supply all the hypotheses of buildings that can be deduced from the arrangement of planes. Among these hypotheses the choice is done through a bayesian formulation that balance data adequacy and caricature needs. Results are provided on real images and show the validity of the approach that remains very generic on the contrary to model-based methods while bringing external architectural information through geometric constraints, which generally lacks in data-driven algorithms.