面向三维点云的域自适应学习

Domain adaptation learning for 3D point clouds:A survey

三维点云数据在自动驾驶、机器人和高精地图等领域得到了广泛应用。目前,基于深度学习的三维点云数据处理主要基于有监督学习,其算法性能依赖于大规模高质量的标注数据集。此外,仅在单一设备与场景中训练的三维点云数据处理模型难以应用于不同设备与环境,泛化性能有限。因此,如何减少三维点云标注数据集的需求以及提高三维点云处理模型的适应性是当前三维点云数据处理面临的重要难题。作为迁移学习的一个重要分支,域自适应学习旨在不同域间特征分布存在差异的情况下提高模型的适应性,可为解决上述难题提供重要思路。为便于对点云域自适应学习领域进行更深入有效的探索,本文主要从对抗学习、跨模态学习、伪标签学习、数据对齐及其他方法5个方面对近年来的三维点云域自适应学习方法进行了系统梳理与分类归纳,并分析总结每类点云域自适应学习方法所具备的优势及面临的问题。最后,对三维点云域自适应学习研究领域的未来发展进行了展望。

Three-dimensional(3D)point cloud data have been widely used in many fields,such as autonomous driving,robotics,and highprecision mapping.At present,the state-of-the-art deep learning-based methods for 3D point cloud processing are mainly supervised learning methods.The performance of these methods depends heavily on large-scale,high-quality annotated datasets.However,annotating a large-scale,high-quality,category-diverse,and scenario-rich dataset is time-consuming and labor-intensive.In particular,obtaining sufficiently large numbers of samples for model optimization is also quite difficult in some special cases.In addition,3D point cloud processing models trained on a single device in a special environment are difficult to generalize to different devices and environments.Their generalizability to various devices and environments is limited.Thus,how to reduce dependencies on high-quality annotated 3D point cloud datasets and how to improve the generalizability of current point cloud processing models are important research topics.In recent years,various kinds of impressive and elaborate technologies,such as meta-learning,few-shot learning,transfer learning,self-supervised learning,semisupervised learning,and weakly supervised learning,have been proposed to solve this problem.As an important research branch of transfer learning,domain adaptive learning aims to eliminate differences in feature distributions across domains and promote the generalization ability of deep learning models,thereby providing a novel solution to address this problem effectively.The academic community has conducted preliminary research on domain adaptive learning for point cloud processing.However,the domain adaptive learning field for point clouds still requires in-depth and effective exploration.Consequently,this study systematically summarizes and classifies recent 3D point cloud domain adaptive learning methods into five categories:adversarial learning,cross-modal learning,pseudolabel learning,data alignment,and other kinds of methods.First,we present the mathematical definition of the domain adaptive learning task and depict the chronological overview of the development of different domain adaptive learning methods to provide readers with a clear understanding.Second,we present the general solution for each category of domain adaptive learning methods and summarize the advantages and disadvantages of the current methods for each category.Third,we compare the performance of current methods on threepoint cloud processing tasks,including 3D shape classification,3D object detection,and 3D semantic segmentation.For each task,we also summarize the commonly used datasets and evaluation metrics for an intuitional comparison.Finally,we conclude the advantages and disadvantages of these five categories of methods and discuss future research directions about the 3D point cloud domain adaptive learning.