Hierarchical shape-based surface reconstruction for dense multi-view stereo

The recent widespread availability of urban imagery has lead to a growing demand for automatic modeling from multiple images. However, modern image-based modeling research has focused either on highly detailed reconstructions of mostly small objects or on human-assisted simplified modeling. This paper presents a novel algorithm which automatically outputs a simplified, segmented model of a scene from a set of calibrated input images, capturing its essential geometric features. Our approach combines three successive steps. First, a dense point cloud is created from sparse depth maps computed from the input images. Then, shapes are robustly extracted from this set of points. Finally, a compact model of the scene is built from a spatial subdivision induced by these structures: this model is a global minimum of an energy accounting for the visibility of the final surface. The effectiveness of our method is demonstrated through several results on both synthetic and real data sets, illustrating the various benefits of our algorithm, its robustness and its relevance for architectural scenes.