A system for performing recognition of three-dimensional objects from photographic images using belief networks has been demonstrated by Mann [54]. Mann has augmented the SUCCESSOR system by adding the capability to create belief networks and to solve them using a software package called HUGIN [5]. This package accepts a belief network as input and generates probabilities of various hypotheses in the network. To demonstrate the feasibility of his approach, Mann has developed a network to recognize the shape and pose of pipe elbows in a greyscale image. Just as in Levitt's system, Mann dynamically creates a belief network using low-level features from an image and groups these features into higher-level objects. In his model, the pipe elbow is made up of several subparts: the female cylinder, the male cylinder, and the spherical joint. Each of these subparts gives rise to certain two-dimensional aggregate features. For instance, the cylindrical parts give rise to parallel ellipses on both ends that are joined by parallel line segments. The aggregated features themselves can be decomposed into various edges and shaded regions. Finally, Mann can reconstruct a model and create a two-dimensional greyscale image of the original pipe elbow.