This thesis introduces a new flexible framework for surfaceconstruction from acquired point sets. This framework casts the surface reconstruction problem as a cells binary labeling problem on a point-guided cell complex under a combination of visibility constraints. This problem can be solved by computing a simple minimum s-t cut allowing an optimal visibility-consistent surface to be efficiently found. In the first part of this thesis, the framework is used for general surface reconstruction problems. A first application leads to an extremely robust surface reconstruction algorithm for dense point clouds from range data. A second application consists in a key component of a dense multi-view stereo reconstruction pipeline, combined with a carefully designed photometric vari- ational refinement. The whole pipeline is suitable to large-scale scenes and achieves state-of-the-art results both in completeness and accuracy of the obtained reconstructions. In the second part of this thesis, the problem of directly reconstructing geometrically simple models from point clouds is addressed. A robust algorithm is proposed to hierarchically cluster a dense point clouds into shapes from a predefined set of classes. If this set of classes is reduced to planes only, the concise reconstruction of models of extremely low combinatorial complexity is achieved. The extension to more general shapes trades this conciseness for a more verbose reconstruction with the added feature of handling more challenging point clouds.
