Unsupervised Domain Adaptation for 3D Point Clouds by Searched Transformations

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