On Recognition of Object Categories

A high level representation of polyhedral scenes in terms of planes and corresponding coplanar sets of 3D line segments is used to develop a method for identifying categories of objects and features in the scenes. Plane intersections are used to establish links between all the planes that correspond to the visible surface of a particular object. The object's shape (as far as it is known) is then reconstructed to provide its description and also constraints on its possible interpretations. At the same time the segment distribution within each plane is analyzed to search for any characteristic patterns that may help identification. In this way we use the topology of a 3D shape or a 2D segment pattern to identify a category of an object (a desk) or a feature (a window), rather that using a metric description of particular object or feature to find its instance(s) in the scene.