We address the reconstruction of CAD models from 3D point clouds, assuming that an a priori CAD model, roughly similar to the scene to reconstruct, is given. The problem can be solved using a Bayesian approach which states the reconstruction task as the search of the most probable CAD model. This article presents a reliable cylinder generation algorithm that can be used to efficiently reach the configuration which maximizes the target probability, combined with an efficient specific optimization heuristic. Results illustrate how our approach can automatically reconstruct CAD models of industrials scenes.