PART SEGMENTATION OF SLICE DATA USING REGULARITY

This paper describes a two-stage technique for inferring part descript ions of articulated objects from parallel planar cross-sections or sli ces. First the data are segmented into significant components or parts , then sticks and blobs are fitted to the parts. We introduce a defini tion of 'a part' in the framework of regularity-based segmentation. Th e segmentation is completely model-independent. It exploits shape regu larities to generate a set of perceptually plausible parts which can b e described by any volumetric primitives. In contrast to several relat ed approaches, bent parts can be successfully analyzed by our system. The volumetric primitives used in the models are meant to be suggestiv e more than literal. An inferred model describes an object in terms of type, connectivity, position and orientation of its parts. Attractive features of the approach include simplicity, good generality and its treatment of bent parts. Experimental results are presented.