Abstract
Deep models have been studied in point cloud classification for the applications of autonomous driving and robotics. One challenging issue is that the point cloud of the same object could be discrepantly captured depending on sensors. Such a difference is the main cause of the domain gap. The deep models trained with one domain of point clouds may not work well with other domains because of such a domain gap. A technique to reduce domain inconsistency is domain adaptation. In this paper, we propose an unsupervised domain adaptation with two novel schemes. First, to improve unreliable pseudo-label assignment, we introduce a voting-based procedure based on the recycling max pooling module, which involves self-paced learning. It helps to increase the training stability of the models. Second, to learn the geometrical characteristics of point clouds in unfamiliar settings, we propose a training method of cutting plane identification, which works in an unsupervised way. Testing with the popular point cloud dataset of PointDA-10 and Sim-to-Real, experiments show that our method increase classification accuracy by 6.5%-points on average, ModelNet and ShapeNet as the source domain and ScanNet, and ScanObjectNN as the target domain. From an ablation study, it was observed that each method contributes to improving the robustness of domain adaptation.