Disentangling Global Orientation with Test-Time Rectification for Unsupervised Non-rigid Point Cloud Correspondence

Unsupervised non-rigid point cloud correspondence aims to identify dense matches between deformable point cloud pairs without ground truth annotations. While existing methods effectively handle shape deformations, they often struggle under complex global orientation changes involving large rotation angles and arbitrary rotation axes. In this paper, we propose a robust test-time point cloud correspondence method that disentangles complex orientation changes from non-rigid shape deformation. We achieve this by modeling global orientation changes as rigid transformations via deformation-invariant priors from pre-trained models to disentangle complex orientation changes while remaining insensitive to shape deformation. Moreover, to resolve symmetry ambiguities induced by complex orientation changes, we augment point clouds with rotational symmetry transformations along learned axes to capture subtle disambiguating cues. To evaluate the impact of arbitrary and complex orientation changes, we construct corrupted versions of existing datasets encompassing varying rotation scales and axes. Extensive experiments reveal the effectiveness of our proposed method.