This paper describes both method and first results of an indoor / outdoor 3D reconstruction application. Images are supposed to come from man-made environments. The method is based on a new cooperative segmentation designed from a rough model of such scenes. This modelisation consists of classically representing scenes by a universe composed of polyhedra, that we call general model. This allows to deduce an operating model, constituted by all the typical 3D tokens of a polyhedron: edges, vertices, plane faces. Their projections are 2D primitives as edge segments, junctions of segments, regions. Then a graph groups these features together . The segmentation principle is based on a loop of two detectors of dual image properties (homogeneity and disparity): each of them is controlled by the results of the other.