We describe an approach that generates a complete b-rep description of a polyhedral object from (geometric primitives derived from) dense range images taken from multiple view-points. Our approach, starting from basic face models of visible surfaces of objects in each local view, matches certain geometric features, extracts rigid-body transformations that relate the local views, and incrementally merges the face models (in local views) into a global 3-dimensional b-rep description of the object. A convenient and effective termination criterion is designed to monitor the merging process. The emphasis is on the use of geometric constraints in building a complete 3-dimensional model of the object.
We have implemented this system in C, running on a SUN Sparcstation. The system, as presented, has been tested on face models derived from several synthetic and real range images and performs successfully with realistic noise levels.