Many tasks in modern urban planning require 3-dimensional (3D) spatial information, preferably in the form of 3D city models. Constructing such models requires automatic methods for reliable 3D building reconstruction. House roofs encountered in residential areas in European cities exhibit a wide variety in their shapes. This limits the use of predefined roof models for their reconstruction. The strategy put forward in this paper is, first, to construct a polyhedral model of the roof structure, which captures the topology of the roof, but which might not be very accurate in a metric sense; and then, in a second step, to improve the metric accuracy by fitting this model to the data. This decoupling of topology extraction from metric reconstruction allows a more efficient roof modelling involving less criteria. And, restricting the processing, at all stages, to one or just a few roof structures, by using a colour-based segmentation of the images, allows to use constraints that are not very tight. The approach has been tested on a state-of-the-art dataset of aerial images of residential areas in Brussels.