SG^2SLAM: Spatial Geometry-Informed 3D Gaussian Splatting SLAM for Monocular Outdoor Navigation

Ziang Guo; Didar Seyidov; Yixiao Zhou; Xuefeng Zhang; Ali Alridha Abdulkarim; Chen Min; Zufeng Zhang; Dzmitry Tsetserukou

cs.ROOct 26, 2026

SG^2SLAM: Spatial Geometry-Informed 3D Gaussian Splatting SLAM for Monocular Outdoor Navigation

Authors: Ziang Guo, Didar Seyidov, Yixiao Zhou, Xuefeng Zhang, Ali Alridha Abdulkarim, Chen Min, Zufeng Zhang, Dzmitry Tsetserukou

Organizations: Intelligent Space Robotics Laboratory, Skolkovo Institute of Science and Technology, Moscow, Russia. · Department of Electrical and Electronic Engineering, The University of Hong Kong, China. · SuZhou Automotive Research Institute, Tsinghua University, China. · Research Center for Intelligent Computing Systems, SKLP, Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China. ·

Abstract

Monocular outdoor navigation requires robust camera motion estimation and a dense, navigable map, where 3D Gaussian Splatting SLAM serves this task through an explicit differentiable parameterization and closed-form projection. However, typically, the geometry is recovered by associating 2D pixels with 3D points, so a modular SLAM system absorbs tracking error into the map, leading to cascading degradation. To address these gaps, we present SG^2SLAM, which couples streaming spatial geometry with a ray-conditioned 3DGS kernel. Streaming spatial geometry replaces the per- frame cross-dimensional correspondence with a chunk-level metric prior that supervises tracking and mapping from one source. A ray-conditioned kernel extends each primitive with a viewing-direction subspace that senses the monocular camera viewpoints. The joint optimization over spatial geometry and poses ensures a globally consistent map and trajectory outcome. Experiments on the KITTI and Waymo Open datasets show that our SG^2SLAM surpasses the per-frame association baselines in the monocular outdoor navigation task. Code will be made publicly available upon acceptance.

Figures & tables

Fig. 1 : $\text{SG}^{2}$ SLAM system. Streaming spatial geometry prior breaks the limitations on a per-frame 2D-to-3D association, optimizing the drift accumulation from cross-dimensional association. 3DGS primitives are ray-conditioned and spatial-geometry-supervised, benefiting the map consistent and geometrically grounded under monocular camera condition. The chunk fusion is solved over the spatial geometry and poses via a global joint graph.

Panel of Figure 2 — Fig. 2 : Qualitative results of 3DGS rendered RGB image on the KITTI dataset. Our $\text{SG}^{2}$ SLAM renders better details both in foreground and background based on our monocular-depth-aware 3DGS kernel.

Panel of Figure 3 — Fig. 3 : Qualitative results of 3DGS rendered depth image on the KITTI dataset. Our rendered depth effectively removes the floating primitives that are harmful to the downstream navigation.

Fig. 4 : Per-frame computational cost of $\text{SG}^{2}$ SLAM.

Methods	Metrics	00	01	02	03	04	05	06	07	08	09	10
	PSNR $\uparrow$	10.09	16.40	8.78	11.83	17.43	10.83	12.66	8.69	6.90	9.71	8.66
MonoGS [ 25 ]	SSIM $\uparrow$	0.43	0.58	0.31	0.36	0.55	0.37	0.43	0.34	0.27	0.27	0.32
	LPIPS $\downarrow$	0.82	0.66	0.85	0.82	0.55	0.84	0.76	0.86	0.89	0.81	0.84
	PSNR $\uparrow$	6.96	13.94	—	14.95	17.90	—	12.68	—	—	11.03	12.82
S3PO-GS [ 30 ]	SSIM $\uparrow$	0.64	0.55	—	0.39	0.54	—	0.44	—	—	0.33	0.41
	LPIPS $\downarrow$	0.90	0.68	—	0.73	0.57	—	0.75	—	—	0.83	0.78

TABLE I : Rendering performance on KITTI [ 35 ] dataset. LC = Loop Closure. OOM = GPU out of memory, — = tracking failure.

Seq.	LC	MonoGS	S3PO-GS	DiskChunGS	GigaSLAM	Ours
00	$\checkmark$	200.92	168.97	OOM	OOM	5.42
01	$\times$	436.78	214.88	208.82	209.17	22.34
02	$\checkmark$	—	—	OOM	OOM	9.83
03	$\times$	25.63	18.75	1.34	1.33	1.77
04	$\times$	18.65	5.99	1.72	1.69	1.65
05	$\checkmark$	—	—	4.85	4.54	5.09

TABLE II : Tracking Performance on KITTI [ 35 ] dataset (ATE RMSE in meters

\downarrow

). LC = Loop Closure. OOM = GPU out of memory, — = tracking failure.

Seq.	MonoGS	S3PO-GS	DiskChunGS	GigaSLAM	Ours
100613	6.96	0.66	0.30	0.36	0.18
106762	18.01	0.61	0.12	0.43	0.20
132384	12.04	1.54	0.05	1.23	0.67
13476	3.30	—	0.45	3.18	0.92
152706	9.30	0.96	0.29	0.63	0.22
153495	5.99	2.03	—	0.73	0.55

TABLE III : Tracking Performance on Waymo Open Dataset [ 36 ] dataset (ATE RMSE in meters

\downarrow

). OOM = GPU out of memory, — = tracking failure.

Panel of Figure 8 — Fig. 5 : Tracking comparison on KITTI dataset and our own sequence without loop closure.

Ablations	ATE RMSE $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Vanilla 3DGS	1.49	22.12	0.78	0.50
Our 3DGS	1.49	23.41	0.89	0.39
Pair-wise Frontend	2.65	19.95	0.67	0.59
Streaming Frontend	1.49	23.41	0.89	0.39
200 frames / chunk	1.49	23.41	0.89	0.39
300 frames / chunk	1.92	22.42	0.80	0.43

TABLE IV : Ablation study on our design for the main modules and the effects of different setups.

3DGS SLAM

3D Gaussian Splatting simultaneous localization and mapping builds renderable 3D Gaussian maps while estimating camera motion online.

See all 63 papers

GaussianFlow SLAM: Monocular Gaussian Splatting SLAM Guided by GaussianFlow
Apr 17, 2026Dong-Uk Seo, Jinwoo Jeon, Eungchang Mason Lee +13D Gaussian Splatting 3DGS SLAM
School of Electrical Engineering, KAIST (Korea Advanced Institute of Science and Technology), Daejeon 34141, Republic of Korea
Robust and Efficient Monocular 3D Gaussian SLAM for Kilometer-Scale Outdoor Scenes
Jun 29, 2026Sicheng Yu, Dongxu Shen, Beizhen Zhao +2Simultaneous Localization and Mapping 3D Gaussian Splatting
The Hong Kong University of Science and Technology (Guangzhou), China
MonoEM-GS: Monocular Expectation-Maximization Gaussian Splatting SLAM
Apr 12, 2026Evgenii Kruzhkov, Sven Behnke3DGS SLAM Camera Pose Estimation
Autonomous Intelligent Systems group, Computer Science Institute VI – Intelligent Systems and Robotics – and the Center for Robotics and the Lamarr Institute for Machine Learning and Artificial Intelligence, University of Bonn, Germany

Apr 17, 2026J/K move · Enter open · S save

GaussianFlow SLAM: Monocular Gaussian Splatting SLAM Guided by GaussianFlow

Dong-Uk Seo, Jinwoo Jeon, Eungchang Mason Lee, Hyun Myung

School of Electrical Engineering, KAIST (Korea Advanced Institute of Science and Technology), Daejeon 34141, Republic of Korea

3D Gaussian Splatting 3DGS SLAM Camera Pose Estimation Visual SLAM Gaussian Splatting

Open paper AskPDF arXiv

Gaussian splatting has recently gained traction as a compelling map representation for SLAM systems, enabling dense and photo-realistic scene modeling. However, its application to monocular SLAM remains challenging due to the lack of reliable geometric cues from monocular input. Without geometric supervision, mapping or tracking could fall in local-minima, resulting in structural degeneracies and inaccuracies. To address this challenge, we propose GaussianFlow SLAM, a monocular 3DGS-SLAM that leverages optical flow as a geometry-aware cue to guide the optimization of both the scene structure and camera poses. By encouraging the projected motion of Gaussians, termed GaussianFlow, to align with the optical flow, our method introduces consistent structural cues to regularize both map reconstruction and pose estimation. Furthermore, we introduce normalized error-based densification and pruning modules to refine inactive and unstable Gaussians, thereby contributing to improved map quality and pose accuracy. Experiments conducted on public datasets demonstrate that our method achieves superior rendering quality and tracking accuracy compared with state-of-the-art algorithms. The source code is available at: https://github.com/url-kaist/gaussianflow-slam.

Abstract

3DGS SLAM

Explore similar work

GaussianFlow SLAM: Monocular Gaussian Splatting SLAM Guided by GaussianFlow

Robust and Efficient Monocular 3D Gaussian SLAM for Kilometer-Scale Outdoor Scenes

MonoEM-GS: Monocular Expectation-Maximization Gaussian Splatting SLAM