Abstract
While techniques that segment shapes into visuallymeaningful parts have generated impressive results, thesetechniques also have only focused on relatively simpleshapes, such as those composed of a single object eitherwithout holes or with few simple holes. In many applica-tions, shapes created from images can contain many over-lapping objects and holes. These holes may come from sen-sor noise, may have important parts of the shape or maybe arbitrarily complex. These complexities that appear inreal-world 2D shapes can pose grand challenges to the ex-isting part segmentation methods. In this paper, we proposea new decomposition method, called Dual-space Decompo-sition that handles complex 2D shapes by recognizing theimportance of holes and classifying holes as either topolog-ical noise or structurally important features. Our methodcreates a nearly convex decomposition of a given shape bysegmenting both the polygon itself and its complementary.We compare our results to segmentation produced by non-expert human subjects. Based on two evaluation methods,we show that this new decomposition method creates statis-tically similar to those produced by human subjects.