G${2}$VLO: Accurate and Generic 2D Gaussian Based Visual-LiDAR Odometry
Diantao Tu, Hainan Cui, Peilin Tao, Yangdong Liu, Shuhan Shen
Abstract
Multimodal SLAM is an important topic in 3D computer vision research. Recent visual-LiDAR SLAM systems use photometric error for camera pose estimation, but their use of sparse LiDAR projections underutilizes image information. Integrating 3D Gaussian Splatting allows full image rendering, but the ellipsoidal Gaussian representation provides an imprecise approximation of scene geometry, introducing bias into geometry-sensitive pose estimation. To address these limitations, this letter presents G<inline-formula><tex-math notation="LaTeX">$^{2}$</tex-math></inline-formula>VLO, an accurate and generic multimodal odometry based on 2D Gaussian Splatting. Compared to 3DGS, 2DGS offers better geometry representation by using 2D Gaussian disks. G<inline-formula><tex-math notation="LaTeX">$^{2}$</tex-math></inline-formula>VLO consists of three components: a LiDAR subsystem, a visual subsystem, and a joint 2DGS mapping subsystem. The LiDAR and visual subsystem process data independently, making G<inline-formula><tex-math notation="LaTeX">$^{2}$</tex-math></inline-formula>VLO a generic framework that supports diverse sampling rates between the LiDAR and camera (5-15 Hz). The joint mapping subsystem uses 2D Gaussians to improve geometry modeling and pose estimation. To further refine camera poses in the visual subsystem, we introduce a photometric bundle adjustment based on the rendered depth and normal. Experiments on public and self-collected datasets show that G<inline-formula><tex-math notation="LaTeX">$^{2}$</tex-math></inline-formula>VLO outperforms state-of-the-art methods.
BibTeX
@inproceedings{ral2026_g2vloaccurateand,
title = {G${2}$VLO: Accurate and Generic 2D Gaussian Based Visual-LiDAR Odometry},
author = {Diantao Tu and Hainan Cui and Peilin Tao and Yangdong Liu and Shuhan Shen},
booktitle = {RA-L 2026},
year = {2026}
}