Learning to Drop Points for LiDAR Scan Synthesis

Abstract

Generative modeling of 3D scenes is a crucial topic for aiding mobile robots to improve unreliable observations. However, despite the rapid progress in the natural image domain, building generative models is still challenging for 3D data, such as point clouds. Most existing studies on point clouds have focused on small and uniform-density data. In contrast, 3D LiDAR point clouds widely used in mobile robots are nontrivial to be handled because of the large number of points and varying-density. To circumvent this issue, 3D-to-2D projected representation such as a cylindrical depth map has been studied in existing LiDAR processing tasks but susceptible to discrete lossy pixels caused by failures of laser reflection. This paper proposes a novel framework based on generative adversarial networks to synthesize realistic LiDAR data as an improved 2D representation. Our generative architectures are designed to learn a distribution of inverse depth maps and simultaneously simulate the lossy pixels, which enables us to decompose an underlying smooth geometry and the corresponding uncertainty of laser reflection. To simulate the lossy pixels, we propose a differentiable framework to learn to produce sample-dependent binary masks using the Gumbel-Sigmoid reparametrization trick. We demonstrate the effectiveness of our approach in synthesis and reconstruction tasks on two LiDAR datasets. We further showcase potential applications by recovering various corruptions in LiDAR data.