This paper describes an algorithm for generating compact feature-based 3D models of indoor environments with a mobile robot. The emphasis lies on the high performance of the algorithm, its possible incremental use, as well as its wide applicability to a variety of sensors as it does not assume any structure in the raw data at all. It recovers planar surfaces of physical environments based on a set of unorganized points and generates a compact, real-time renderable 3D model.