Learning Cross-Domain Features for Domain Generalization on Point Clouds

Abstract

AbstractModern deep neural networks trained on a set of source domains are generally difficult to perform well on an unseen target domain with different data statistics. Domain generalization (DG) aims to learn a generalized model that performs well on the unseen target domain. Currently, most DG approaches are applied to images, and there is less related research in the field of point cloud. In this paper, we propose a novel cross-domain feature learning network architecture for DG on 3D object point clouds, which learns domain invariant representation via data augmentation and hierarchical features alignment (HFA). The data augmentation is empowered by two subtasks: (1) A point set mask on source data such that some parts of the point cloud are removed randomly, to capture domain-shared representation of semantic categories; (2) A linear mixup of different source domain point cloud samples, to address the large domain gap between different domains. HFA is used to align multi-level local features and narrow the distribution distance between different domains. Since there is no common evaluation benchmark for 3D point cloud DG scenario, we experiment on the PointDA-10 and PointSegDA datasets, and extend point cloud domain adaptation (DA) methods to DG for comparison. Our method exhibits superiority in classification and segmentation accuracy over state-of-the-art general-purpose DA methods.KeywordsDomain generalizationDomain-invariant representationHierarchical feature alignment