This paper presents a method for localizing polyhedra using CAD data of object and range images. It is assumed that the type of object is known and that it lies in a stable position on a stand without any overlap. Before image processing geometric/local constraints and stable attitude/global constraints of objects are extracted from CAD data. From the input image a shape map and jump and roof-edge maps based on curvature are extracted. These maps are used for object decomposition under the assumption of: (1) edges partition faces, and (2) shape boundaries partition faces.

From partitioned faces a description of an object based on its attributes and connectivity relations with other faces is given. Matching is performed between this description and a set of models. Constraints obtained from CAD data limit considerably candidates for correspondence thus making efficient matching possible.

Finally, we verify the realizability of obtained matching results and determine the attitude. To determine the attitude of an object, it suffices to determine three degrees of freedom. This paper describes not only the construction and methodology of the system but also shows experimental results based on real data.