Improving Sparse Point Cloud Classification Performance Through Image Inpainting

With the increasing number of platforms using LiDAR, robots using LiDAR have become more easily accessible in our daily lives. LiDAR sensors are essential sensors for robots, much like eyes, but their limitations are also evident. Due to the nature of LiDAR sensors, the amount of points that can be obtained from distant objects is sparse, which has a significant negative impact on the object classification performance through LiDAR. This paper proposes augmenting the point cloud through image inpainting to overcome this limitation. The point cloud collected through LiDAR is transformed into a 2D image, restored through image inpainting, and then validated for performance using the Point-wise Classification technique.