Learnable Bi-directional Data Augmentation for few-shot cross-domain point cloud classification

Xu Sun, Lin Zhang, Jiahao Xu, Jiakang Yuan, Tao Chen

Abstract: This paper studies the 3D Few-Shot Domain Adaptation (FSDA) task, which aims to adapt models from one domain to another using only a few labeled target point clouds. Compared with the widely adopted Unsupervised Domain Adaptation (UDA) problem setting, which assumes a large number of unlabeled target data is available, obtaining few-shot labeled target data costs less. For 3D point clouds, local geometric structure information is important in domain adaptation, making previous high-level global feature alignment-based methods used in 2D FSDA no longer suitable. Therefore, we propose to use point cloud augmentation to enhance geometric diversity and design a Learnable Bi-directional Data Augmentation (LBDA) approach to tackle the 3D FSDA task. Our proposed LBDA enhances source and target domain data, transforming them into easily transferable representations that bridge the domain gap by a proposed ellipsoid-based loss constraint. Additionally, we present a Sample-level Adaptive Weighting (SAW) method, allowing the DA model to selectively focus more on readily transferable samples and vice versa, thus mitigating the risk of negative transfer. Our experimental results on the PointDA-10 dataset and Sim-to-Real dataset showcase the effectiveness of our approach in the 5-shot scenario, outperforming state-of-the-art UDA methods. This paper contributes to the burgeoning field of point cloud analysis, offering a promising solution to the 3D Few-Shot Domain Adaptation problem.