Abstract
Large garages are ubiquitous yet intricate scenes in our daily lives, posingchallenges characterized by monotonous colors, repetitive patterns, reflectivesurfaces, and transparent vehicle glass. Conventional Structure from Motion(SfM) methods for camera pose estimation and 3D reconstruction fail in theseenvironments due to poor correspondence construction. To address thesechallenges, this paper introduces LetsGo, a LiDAR-assisted Gaussian splattingapproach for large-scale garage modeling and rendering. We develop a handheldscanner, Polar, equipped with IMU, LiDAR, and a fisheye camera, to facilitateaccurate LiDAR and image data scanning. With this Polar device, we present aGarageWorld dataset consisting of five expansive garage scenes with diversegeometric structures and will release the dataset to the community for furtherresearch. We demonstrate that the collected LiDAR point cloud by the Polardevice enhances a suite of 3D Gaussian splatting algorithms for garage scenemodeling and rendering. We also propose a novel depth regularizer for 3DGaussian splatting algorithm training, effectively eliminating floatingartifacts in rendered images, and a lightweight Level of Detail (LOD) Gaussianrenderer for real-time viewing on web-based devices. Additionally, we explore ahybrid representation that combines the advantages of traditional mesh indepicting simple geometry and colors (e.g., walls and the ground) with modern3D Gaussian representations capturing complex details and high-frequencytextures. This strategy achieves an optimal balance between memory performanceand rendering quality. Experimental results on our dataset, along withScanNet++ and KITTI-360, demonstrate the superiority of our method in renderingquality and resource efficiency.