LiDAR Data Augmentation for Semantic Segmentation: A Statistical Intensity Model

Light Detection and Ranging (LiDAR) sensors play a crucial role in autonomous driving, specially in HDMaps generation systems, which contain semantic information of the road scenarios.Gathering this information is not easy, so in the literature deep learningbased semantic segmentation on LiDAR point clouds is used to accelerate this process.However, it requires large amounts of data for training, which are time-consuming and computationally intensive to obtain.In this paper, we introduce a method to use the synthetic dataset KITTI-CARLA to reduce the amount of real data needed to train a Deep Learning-based model on SemanticKITTI.The main challenge is that in KITTI-CARLA the intensity field is not computed.To overcome this, we define a statistic module that infer the real distribution of intensities of SemanticKITTI into KITTI-CARLA.We tested our method with the Point-to-Voxel-KD model.Extensive experiments show that results near the state of the art can be obtained with half the amount of real data when KITTI-CARLA is incorporated to the training set.The source code for the intensity distribution calculation and the conversion process from KITTI-CARLA to SemanticKITTI is published in https://github.com/pabloaguilarp/KITTI-CARLA-converter.