Havaldar and Medioni describe a passive vision system for the recovery of volumetric object descriptions from a trinocular imaging system. Edges obtained from each of the views are matched interocularly to obtain contours. These contours are then labeled as either true edges (corresponding to 3D surfaces) or limb edges (corresponding to smooth generalized cylinders) and used to generate a generalized cylinder representation of the object in view. A correlation-based stereopsis algorithm is then applied near the represented surfaces in order to deal with local deformations.
The six pages allocated to this paper are insufficient to describe the algorithm fully. To obtain a more complete description of the algorithm and its performance, one must refer to earlier conference proceedings, which are cited in the text.