Existing tools for 3D building reconstruction usually apply approaches, which are either based on constructive solid geometry (CSG) or boundary representation (B-Rep). After a brief discussion of their respective advantages and disadvantages, the paper will present an alternative approach based on cell decomposition. This type of representation is also well known in solid modelling, and can be used efficiently for building reconstruction. Firstly, topological correct representations of building polyhedrons can be constructed easily from planar surface patches as they can for example be extracted from aerial LIDAR data. Furthermore, constraints between different object parts like co-planarity or right angles can be integrated relatively easy. The approach will be demonstrated exemplarily by a building reconstruction based on aerial LIDAR data and given outlines of the respective buildings. In principle, different levels of generalisation can be defined during reconstruction. This also allows a refinement of an already given building model based on terrestrial LIDAR data as it will be demonstrated in the final part of the paper.