Chen addresses the problem of identifying the presence, orientation, and location of a possibly occluded 3D rigid object, with arbitrary curved surfaces, from a single stationary 2D image. A model-based recognition framework is adopted. An object model is constructed from a central image plus four other training images, taken from neighboring views that define a model aspect. Thus an aspect is a function of the viewpoint, and several aspects may be needed to cover the interesting views of an object. Some image processing steps are needed for the detection of the object’s silhouette (in the central image) as well as the internal edges (caused by discontinuities in the object surface or by shadows) visible within the aspect being modeled. The model data include the surface curvature at each silhouette point, in addition to view-dependent 3D information at each silhouette and internal edge point. This modeling technique allows the generation of object appearance for viewpoints within a model aspect using a transformation of coordinates with six degrees of freedom.

The contribution of the paper is in the formulation of the matching problem as a least squares minimization problem of a given distance function between certain points in both observed and model data. The Newton iterative technique is used to adjust the pose parameters so that the model’s appearance gets closer to the observed data. Heuristics are given to speed the matching process when occlusion is present. Some results obtained on a Sun workstation are given. The paper reflects the progress in viewer-centered (as opposed to object-centered) 3D shape recognition using visual information.