LI-GS: Gaussian Splatting with LiDAR Incorporated for Accurate Large-Scale Reconstruction

TL;DR

LI-GS tackles the challenge of geometrically accurate large-scale 3D reconstruction by integrating LiDAR with Gaussian Splatting through plane-constrained incremental 4D GMMs. The approach combines LiDAR-guided initialization, a normalization-driven optimization (with L_GMM, photometric, sky, depth, and normal losses), and geometry-aware density control to produce coherent Gaussian surfels and high-quality meshes via a coarse-to-fine, GMM-informed meshing pipeline. Experiments on outdoor and indoor scenes demonstrate state-of-the-art geometric accuracy and competitive rendering quality, with substantial improvements over both LiDAR-only and Gaussian-based baselines. The work provides a practical framework for LiDAR-augmented explicit surface representations, with potential extensions to SLAM and real-time mapping in unbounded environments.

Abstract

Large-scale 3D reconstruction is critical in the field of robotics, and the potential of 3D Gaussian Splatting (3DGS) for achieving accurate object-level reconstruction has been demonstrated. However, ensuring geometric accuracy in outdoor and unbounded scenes remains a significant challenge. This study introduces LI-GS, a reconstruction system that incorporates LiDAR and Gaussian Splatting to enhance geometric accuracy in large-scale scenes. 2D Gaussain surfels are employed as the map representation to enhance surface alignment. Additionally, a novel modeling method is proposed to convert LiDAR point clouds to plane-constrained multimodal Gaussian Mixture Models (GMMs). The GMMs are utilized during both initialization and optimization stages to ensure sufficient and continuous supervision over the entire scene while mitigating the risk of over-fitting. Furthermore, GMMs are employed in mesh extraction to eliminate artifacts and improve the overall geometric quality. Experiments demonstrate that our method outperforms state-of-the-art methods in large-scale 3D reconstruction, achieving higher accuracy compared to both LiDAR-based methods and Gaussian-based methods with improvements of 52.6% and 68.7%, respectively.

Figures (11)

• Figure 1: The performance of LI-GS. (a) The comprehensive mesh of a campus scene, with its details shown in (b)-(e). (f) Our data collection platform. (g)-(h) The rendered RGB image and the colorized mesh of an indoor scene.
• Figure 2: The overview of our system. Our system involves three steps: (a) We utilize the plane-constrained incremental 4D GMM to generate initial Gaussian surfels. (b) To optimize the Gaussian surfels, we apply global image normalization, local GMM normalization, and geometry-aware density control. (c) We eliminate incorrect samples using a coarse-to-fine method, and then apply the screened Poisson reconstruction method kazhdan2013screened to extract meshes.
• Figure 3: Comparison of colorized point clouds generated using two different methods, (a) interpolated image poses and (b) our previous work jiang2024er. Our method produces colorized point clouds with more distinct textures.
• Figure 4: The pipeline of plane-constrained incremental 4D GMM modeling.
• Figure 5: The performance of plane-constrained 4D GMM modeling. GMM components are represented as ellipsoids. (a) Illustration of the input colorized point cloud. (b), (d) The plane-constrained 4D GMMs in two specific areas. (c), (e) The 3D GMMs without plane constraints.