Self-Supervised Boundary Point Prediction Task for Point Cloud Domain Adaptation

Abstract

Unsupervised domain adaptation (UDA) could significantly improve the cross-domain performance of current supervised 3D deep learning methods and have a widespread application prospect. However, the domain gap between source domain and target domain renders the UDA problem highly challenging. In this letter, we present a novel UDA method for point clouds from the perspective of multi-strategy. First, we explore the effectiveness of state-of-the-art data augmentation methods to point cloud domain adaptation, and introduce a data augmentation procedure to two widely-existed scenarios, i.e., sim-to-sim and sim-to-real. And then, we explore a mask deformation procedure to simulate the missing parts with respect to the real-world point clouds. On one hand, the masked point clouds push network to pay more attention to local features rather than global features; on other hand, we employ a prediction-consistency contrastive loss to improve the prediction robustness of network based on the mask deformation. Moreover, we propose a self-supervised learning task by predicting the boundary points of masked region. Specifically, the network could effectively perceive the occlusion and capture fine-grained features by automatically labeling and predicting the boundary points of the marked region. Extensive experiments conducted on both PointDA-10 and PointSegDA benchmarks for point cloud classification and segmentation, respectively, demonstrate the effectiveness of the proposed method.