Splat-LOAM: Gaussian Splatting LiDAR Odometry and Mapping

TL;DR

Splat-LOAM introduces the first LiDAR odometry and mapping pipeline that relies exclusively on 2D Gaussian primitives to represent and refine the scene. By projecting LiDAR data onto a spherical domain and employing a differentiable, tile-based Gaussian rasterizer, the method jointly optimizes local Gaussian maps and aligns new frames through both geometric and photometric cues, yielding competitive odometry and state-of-the-art-like mapping efficiency with reduced GPU requirements. The approach is evaluated on multiple public datasets, demonstrating robust performance, favorable memory/runtime trade-offs, and detailed ablations that highlight the benefits of joint optimization and the Gaussian representation. The work points to practical impact in real-time robotics where lightweight, dense, and consistent scene representations are advantageous, while outlining future work to address motion distortion and loop closure integration.

Abstract

LiDARs provide accurate geometric measurements, making them valuable for ego-motion estimation and reconstruction tasks. Although its success, managing an accurate and lightweight representation of the environment still poses challenges. Both classic and NeRF-based solutions have to trade off accuracy over memory and processing times. In this work, we build on recent advancements in Gaussian Splatting methods to develop a novel LiDAR odometry and mapping pipeline that exclusively relies on Gaussian primitives for its scene representation. Leveraging spherical projection, we drive the refinement of the primitives uniquely from LiDAR measurements. Experiments show that our approach matches the current registration performance, while achieving SOTA results for mapping tasks with minimal GPU requirements. This efficiency makes it a strong candidate for further exploration and potential adoption in real-time robotics estimation tasks.

Figures (10)

• Figure 1: Performance overview of Splat-LOAM. F1 score to Active Memory [Mb] and Runtime [s]. The plots provide a quantitative comparison between state-of-the-art mapping pipelines, while PIN-SLAM and ours also perform online odometry.
• Figure 2: Splat-LOAM Overview. Given a LiDAR point cloud, we leverage the spherical projection to generate an image-like representation. Moreover, using an ad-hoc differentiable rasterizer, we guide the optimization for structural parameters of 2D Gaussians. The underlying representation is concurrently used to incrementally register new measurements.
• Figure 3: Bounding box computation for near-singularity splats.(a) shows the 3D configuration of a splat that approximately lies behind the camera. (b) shows the corresponding spherical image with the projected bounding box vertices. The distortion is removed by shifting the vertices along the horizontal direction to align the projected center to the image center. (c) being far from the coordinate singularity, we compute the maximum extent of the splat. (d) we revert the shift and propagate to the corresponding tiles via an offset from the central vertex, matching with the tiles highlighted on (a).
• Figure 4: RPE evaluation. Number of successful sequences across RPE thresholds. It includes the sequences of Newer College zhang2014loam, VBR bg2024vbr, Oxford Spires tao2024oxspires and Mai City vizzo2021icra.
• Figure 5: Memory and Runtime Analysis. The plot relates the used GPU Memory and the mapping iteration frequency with the number of active primitives. The measurements were obtained over the longest sequence we reported: campus bg2024vbr.