We present a fully automatic pipeline for modeling buildings in an urban environment from 3D point clouds. In contrast to existing approaches, our pipeline is robust, scalable and provides a more complete description by not making a Manhattan-World assumption and modeling both buildings (with polyhedra) as well as the non-planar ground, while maintaining the natural inherent smoothness of the ground. It relies on probability theory and scales to datasets with hundreds of millions of noisy 3D points derived from images or LiDAR scans. The pipeline is based on an accurate large-scale segmentation method using a combination of RANSAC (RANdom SAmple Consensus) and nonparametric Bayesian clustering, a set of basic rules enabling semantic scene interpretation and detection of shape priors for buildings, polygon-sweeping and template as well as surface fitting for modeling both natural (the ground) and man-made surfaces (buildings). Besides being robust against substantial noise, efficient and scalable, our pipeline operates fully automatically and has very few parameters to tune, hence, enabling an easy usage. We demonstrate the effectiveness of our pipeline on a wide variety of datasets from both the research community as well as our acquisitions.    «

We present a fully automatic pipeline for modeling buildings in an urban environment from 3D point clouds. In contrast to existing approaches, our pipeline is robust, scalable and provides a more complete description by not making a Manhattan-World assumption and modeling both buildings (with polyhedra) as well as the non-planar ground, while maintaining the natural inherent smoothness of the ground. It relies on probability theory and scales to datasets with hundreds of millions of noisy 3D poi...    »