Cross-Domain Point Cloud Recognition with Deep Learning

Abstract

Point cloud recognition using deep learning methods has attracted increasing research interest recently due to its great potential in real-world applications such as autonomous driving, robotics, etc. However, point clouds of similar objects often exhibit notable geometric variations due to the difference in capturing devices or environmental changes. This leads to significant performance degradation when a learnt point cloud recognition model is applied to a new scenario, which is also known as the domain adaptation issue.

In this thesis, we first provide a comprehensive literature review of deep learning on visual recognition, unsupervised domain adaptation, open-set unsupervised domain adaptation, self-supervised learning and knowledge transfer to introduce the background of the thesis. Then, an introduction to the problem setting and the commonly used benchmark datasets is provided for a better understanding of the task. Next, a point-level domain adaptive point sampling (DAPS) strategy is proposed to tackle the domain gap in cross-domain point cloud recognition. In addition, an instance-level domain adaptive cloud sampling (DACS) strategy is proposed to learn additional target-specific information for better recognition performance on the target domain. Moreover, we further propose a two-stage open-set domain adaptive sampling (OS-DAS) strategy to learn an open-set recognition model in a coarse-to-fine manner to tackle the open-set unsupervised domain adaptation issue. Finally, we list some potential research directions for cross-domain point cloud recognition.