Splat-LOAM: Gaussian Splatting LiDAR Odometry and Mapping

Published in IEEE/CVF International Conference on Computer Vision (ICCV), 2025

Emanuele Giacomini, Luca Di Giammarino, Lorenzo De Rebotti, Giorgio Grisetti, Martin R. Oswald

Teaser Image

Abstract

LiDARs provide accurate geometric measurements, making them valuable for ego-motion estimation and reconstruction tasks. However, managing an accurate and lightweight environment representation remains challenging. This work pioneers the use of Gaussian primitives for a LiDAR odometry and mapping pipeline, using spherical projection to refine scene representation from LiDAR data. The approach achieves state-of-the-art mapping performance with minimal GPU requirements, while matching current standards in registration accuracy. Efficiency and effectiveness make it a promising step toward real-time SLAM solutions.

Resources

[arxiv]

Bibtex

```bibtex @inproceedings{giacomini2025splatloam, title={Splat-LOAM: Gaussian Splatting LiDAR Odometry and Mapping}, author={Giacomini, Emanuele and Di Giammarino, Luca and De Rebotti, Lorenzo and Grisetti, Giorgio and Oswald, Martin R.}, booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)}, year={2025} }