Abstract: | We present a system for recovering 3-D convex sub-parts and their connectivity from parallel slices of a range image. We call this the volumetric segmentation problem. Instead of fitting a set of 3-D models to the data, we detect regularities (transversality) in the data in order to find plausible joints between subparts. This defines implicitly the class of objects which can be successfully processed. The system can recover strongly bent subparts even though it analyses only parallel cross-sections. Some experimental results are discussed and evaluated according to two criteria, stability and believability.
|