SplatAD：面向自动驾驶的3D高斯溅写实雷达和摄像头渲染
摘要
确保自动化机器人(如自动驾驶车辆)的安全，需要跨越 diverse driving scenarios 进行大规模 testing。Simulation 是实现 cost-effective and scalable testing 的关键要素。Neural rendering 方法日益流行，能够从收集的日志中 data-driven manner 构建 simulation 环境。然而，现有的 neural radiance field (NeRF) 方法 for sensor-realistic rendering of camera and lidar data suffers from low rendering speeds，限制了其 large-scale testing 的适用性。虽然 3D Gaussian Splatting (3DGS) enables real-time rendering，但 current methods limited to camera data，无法渲染自动驾驶必不可少的 lidar data。为此，我们提出了 SplatAD，首个基于 3DGS 的方法，用于 both camera and lidar data 的 realistic, real-time rendering of dynamic scenes。SplatAD accurately models key sensor-specific phenomena such as rolling shutter effects, lidar intensity, and lidar ray dropouts， using purpose-built algorithms to optimize rendering efficiency。在三个 autonomous driving 数据集上进行评估显示，SplatAD 达到了 state-of-the-art rendering quality， NVS 方面 PSNR 提升最高达 +2， reconstruction 方面 PSNR 提升最高达 +3，而 rendering speed 相对于 NeRF-based methods 提升了一个数量级。
引用
@article{arxiv.2411.16816,
title = {SplatAD: Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving},
author = {Georg Hess and Carl Lindström and Maryam Fatemi and Christoffer Petersson and Lennart Svensson},
journal= {arXiv preprint arXiv:2411.16816},
year = {2025}
}