I. Introduction
3D point clouds serve as concise and versatile representations, providing abundant geometric, shape, and scale details, making them a popular choice for 3D data representation. Training of deep neural networks is typically reliant on large-scale annotated datasets. However, gathering such annotations of 3D point clouds can be laborious and time-consuming due to challenges like occlusion and irregular structure of point clouds. To mitigate this issue, self-supervised learning stands out as a prominent solution.