As Symmetric As Possible: Shape Completion with Non-Rigid Registration Leveraging Generalized Cylinder Decomposition

Abstract

To infer a 3D entire shape from its partial observation, a non-rigid registration algorithm that employs embedded deformations is proposed. We construct a deformation graph on a reference model to discretize the space, and compute a complex deformation as a collection of affine transformations to align the reference model toward the given geometric data. To avoid distortion artifacts during the non-rigid registration, we introduce constraint “As symmetric as possible (ASAP)” on the graph via a generalized cylinder decomposition. ASAP allows model deformation maintaining its underlying local symmetry, which leads to plausible shape completion in the area with no observation. We performed experiments with synthesized data, and demonstrated that the proposed method successfully restored missing surfaces compared with conventional completion techniques.