High-quality data are of utmost importance for any deep-learning application. However, acquiring such data and their annotation is challenging. This paper presents a GPU-accelerated simulator that enables the generation of high-quality, perfectly labelled data for any Time-of-Flight sensor, including LiDAR. Our approach optimally exploits the 3D graphics pipeline of the GPU, significantly decreasing data generation time while preserving compatibility with all real-time rendering engines. The presented algorithms are generic and allow users to perfectly mimic the unique sampling pattern of any such sensor. To validate our simulator, two neural networks are trained for denoising and semantic segmentation. To bridge the gap between reality and simulation, a novel loss function is introduced that requires only a small set of partially annotated real data. It enables the learning of classes for which no labels are provided in the real data, hence dramatically reducing annotation efforts. With this work, we hope to provide means for alleviating the data acquisition problem that is pertinent to deep-learning applications.
Denis, L, Royen, RD, Vercheval, N, Pizurica, A & Munteanu, A 2023, 'GPU Rasterization-Based 3D LiDAR Simulation for Deep Learning', Sensors, vol. 23, no. 19, 8130. https://doi.org/10.3390/s23198130
Denis, L., Royen, R. D., Vercheval, N., Pizurica, A., & Munteanu, A. (2023). GPU Rasterization-Based 3D LiDAR Simulation for Deep Learning. Sensors, 23(19), Article 8130. https://doi.org/10.3390/s23198130
@article{66415a7cbb244d14a58c22d710f37574,
title = "GPU Rasterization-Based 3D LiDAR Simulation for Deep Learning",
abstract = "High-quality data are of utmost importance for any deep-learning application. However, acquiring such data and their annotation is challenging. This paper presents a GPU-accelerated simulator that enables the generation of high-quality, perfectly labelled data for any Time-of-Flight sensor, including LiDAR. Our approach optimally exploits the 3D graphics pipeline of the GPU, significantly decreasing data generation time while preserving compatibility with all real-time rendering engines. The presented algorithms are generic and allow users to perfectly mimic the unique sampling pattern of any such sensor. To validate our simulator, two neural networks are trained for denoising and semantic segmentation. To bridge the gap between reality and simulation, a novel loss function is introduced that requires only a small set of partially annotated real data. It enables the learning of classes for which no labels are provided in the real data, hence dramatically reducing annotation efforts. With this work, we hope to provide means for alleviating the data acquisition problem that is pertinent to deep-learning applications.",
author = "Leon Denis and Royen, {Remco Donovan} and Nicolas Vercheval and Aleksandra Pizurica and Adrian Munteanu",
note = "Funding Information: This research was funded by the Fonds Wetenschappelijk Onderzoek (FWO) (projects G094122N, FWOSB88 - PhD fellowship R. Royen). ",
year = "2023",
month = sep,
day = "28",
doi = "10.3390/s23198130",
language = "English",
volume = "23",
journal = "Sensors",
issn = "1424-8220",
publisher = "Multidisciplinary Digital Publishing Institute (MDPI)",
number = "19",
}