Unsupervised Domain Adaptation for 3D point clouds by searched transformation

Abstract

Input level domain adaptation reduces the burden of a neural encoder by reducing the domain gap at the input level unsupervisedly. It is widely used in 2D visual domain, e.g. , images and videos, but not for 3D point cloud. We first propose to use the input level do- main adaptation for 3D point cloud, namely Point-level domain adaptation. Specifically, we propose to learn a transformation of 3D point clouds by searching the best combination of operations on point clouds which transfer data from source domain to target one while main- taining the classification label without supervision of target label. We decompose the learning objective into two terms, resembling domain shift and preserving label information. On the PointDA-10 benchmark dataset, our method outperforms state-of-the-art unsupervised point cloud domain adaptation methods by large margins (up to + 3.97 % in average).