Learning to Predict Lidar Intensities

We propose a data-driven method for simulating lidar sensors. The reads computer-generated data, and (i) extracts geometrically simulated point clouds (ii) predicts the strength of response – lidar intensities . Qualitative evaluation proposed pipeline demonstrates ability to predict systematic failures such as no/low responses on polished parts car bodyworks windows, or strong reflective surfaces traffic signs license/registration plates. also experimentally show that enhancing training set by data improves segmentation accuracy real dataset with limited access data. Implementation resulting simulator GTA V game, well accompanying large dataset, is made publicly available.