Abstract
Man-made 3D shapes are naturally organized in parts and hierarchies; suchstructures provide important constraints for shape reconstruction andgeneration. Modeling shape structures is difficult, because there can bemultiple hierarchies for a given shape, causing ambiguity, and across differentcategories the shape structures are correlated with semantics, limitinggeneralization. We present StructRe, a structure rewriting system, as a novelapproach to structured shape modeling. Given a 3D object represented by pointsand components, StructRe can rewrite it upward into more concise structures, ordownward into more detailed structures; by iterating the rewriting process,hierarchies are obtained. Such a localized rewriting process enablesprobabilistic modeling of ambiguous structures and robust generalization acrossobject categories. We train StructRe on PartNet data and show itsgeneralization to cross-category and multiple object hierarchies, and test itsextension to ShapeNet. We also demonstrate the benefits of probabilistic andgeneralizable structure modeling for shape reconstruction, generation andediting tasks.