Performance of LiDAR object detection deep learning architectures based on artificially generated point cloud data from CARLA simulator

Training deep neural network algorithms for LiDAR based object detection for autonomous cars requires huge amount of labeled data. Both data collection and labeling requires a lot of effort, money and time. Therefore, the use of simulation software for virtual data generation environments is gaining wide interest from both researchers and engineers. The big question remains how well artificially generated data resembles the data gathered by real sensors and how the differences affects the final algorithms performance. The article is trying to make a quantitative answer to the above question. Selected state-of-the-art algorithms for LiDAR point cloud object detection were trained on both real and artificially generated data sets. Their performance on different test sets were evaluated. The main focus was to determinate how well artificially trained networks perform on real data and if combined train sets can achieve better results overall.