Building reconstruction from lidar data offers promising prospects for rapid generation of large-scale 3D models autonomously. Such reconstruction requires knowledge on a variety of parameters that refer to both the point cloud and the modeled buildings. The complexity of the reconstruction task has led researchers to use external information to localize buildings and assume that they consist of only planar parts. These assumptions limit the reconstruction of complex buildings, particularly those having curved faces. We present in this paper a detection and reconstruction model that considers the point cloud as the only information source and supports the reconstruction of general shape surfaces. Nonetheless, since many of the buildings are composed of planar faces, we maintain the planar based partitioning whenever possible and model non-planar surfaces only where needed. This way, standard models are extended to support free-form roof shapes without imposing artificial models. In addition to the free-form surface extension, we demonstrate the effect of imposing geometric constraints on the reconstruction as a means to generate realistic building models.