Rotational symmetry detection in 3D using reflectional symmetry candidates and quaternion-based rotation parameterization (Q2082604)

The authors present a concept for the detection of rotational symmetry in point clouds or objects that can easily be converted to point clouds. Based on the well-known decomposition of rotations into two reflections, they suggest to use some algorithm for reflection symmetry detection and then combine pairs of symmetry planes to create candidates for rotational symmetry. Among the candidates, maximizers of a specially crafted differentiable symmetry measure are selected. The same symmetry measure is then used to further improve the thus obtained candidates. Attention is paid to make this procedure robust with respect to noisy or incomplete data and to handle approximate symmetry as well. The local optimization allows to detect rotational symmetry even in objects that exhibit only a rather weak reflection symmetry. Rotational symmetries with small revolute angle need to be penalized in order to prevent convergence to the identity but can later be recovered from rotations with suitable integer multiples as revolute angles. For objects with circular symmetry, a more or less random revolute angle will be returned, but circular symmetry can be detected by an a-posteriori test. Systematic and fair comparison to existing algorithms is difficult due to lack of algorithms, implementations, and data sets. The proposed algorithm seems to be general, robust and versatile in the sense that it can easily be adapted to use scenarios with diverse requirements.