Unsupervised Domain Adaptation on Point Clouds via High-Order Geometric Structure Modeling

Abstract

Point clouds can capture the precise geometric information of objects and scenes, which are an important source of 3-D data and one of the most popular 3-D geometric data structures for cognitions in many real-world applications like automatic driving and remote sensing. However, due to the influence of sensors and varieties of objects, the point clouds obtained by different devices may suffer obvious geometric changes, resulting in domain gaps that are prone to the neural networks trained in one domain failing to preserve the performance in other domains. To alleviate the above problem, this article proposes an unsupervised domain adaptation framework, named HO-GSM, as the first attempt to model high-order geometric structures of point clouds. First, we construct multiple self-supervised tasks to learn the invariant semantic and geometric features of the source and target domains, especially to capture the feature invariance of high-order geometric structures of point clouds. Second, the discriminative feature space of target domain is acquired by using contrastive learning to refine domain alignment to specific class level. Experiments on the PointDA-10 and GraspNetPC-10 collection of datasets show that the proposed HO-GSM can significantly outperform the state-of-the-art counterparts.