Unsupervised joint learning of shape and correspondence for non-rigid point cloud registration

Abstract

In recent years, the field of non-rigid point cloud registration has witnessed significant advancements, primarily through the adoption of state-of-the-art, learning-based strategies that revolve around the matching algorithm. However, these deep matching-based methods, which include both soft and hard variants, have been impeded by a notable challenge: the occurrence of imperfect matching. This issue largely stems from disparities in point density between the compared point clouds. To tackle this problem, our method introduces a groundbreaking approach to non-rigid point cloud registration. Our method diverges from traditional practices by simultaneously learning the shape and correspondence of point clouds in an unsupervised manner. Contrary to existing matching-based techniques that attempt to identify direct matches in the target point cloud, our strategy focuses on reconstructing the shape of the target cloud. It then innovatively generates points that align with this understood shape, ensuring a more accurate and balanced registration process. We show the effectiveness of our method through comprehensive experiments on various registration benchmarks, registration task settings, and prominent backbones, yielding unprecedented performance improvement even in the occurrence of imperfect matching.