Large-Scale 3D Point Cloud Representations via Graph Inception Networks with Applications to Autonomous Driving

With the growth of 3D sensing technologies, one can now use a large number of 3D points to precisely represent objects’ surfaces and surrounding environments. We call those 3D points a 3D point cloud; it has a growing impact on various applications, including autonomous driving, drones, robotics, virtual reality and preservation of historical artifacts [4]. For example, a self-driving car could use multiple sensors to observe the world, such as LiDARs, cameras and RADARs [2]. Among those, LiDARs produce two types of 3D point clouds: real-time LiDAR sweeps and high-definition maps. Both types of 3D point clouds provides accurate range information for self-driving cars, which are critical to localization and perception systems. We consider these point clouds large-scale point clouds because they contain a large number of 3D points and record outdoor, open areas.