Learning-based reflection-aware virtual point removal for large-scale 3D point clouds

Abstract

3D point clouds are widely used for robot perception and navigation. LiDAR sensors can provide large scale 3D point clouds (LS3DPC) with a certain level of accuracy in common environment. However, they often generate virtual points as reflection artifacts associated with reflective surfaces like glass planes, which may degrade the performance of various robot applications. In this letter, we propose a novel learning-based framework to remove such virtual points from LS3DPCs. We first project 3D point clouds onto 2D image domain to investigate the distribution of the LiDAR's echo pulses, which is then used as an input to the glass probability estimation network. Moreover, the 3D feature similarity estimation network exploits the deep features to compare the symmetry and geometric similarity between real and virtual points with respect to the estimated glass plane. We provide a LS3DPC dataset with synthetically generated reflection artifacts to train the proposed network. Experimental results show that the proposed method achieves the better performance qualitatively and quantitatively compared with the existing state-of-the-art methods of 3D reflection removal.