VBGS-SLAM: Variational Bayesian Gaussian Splatting Simultaneous Localization and Mapping

Yuhan Zhu; Yanyu Zhang; Jie Xu; Wei Ren

cs.CVApr 3, 2026

VBGS-SLAM: Variational Bayesian Gaussian Splatting Simultaneous Localization and Mapping

Authors: Yuhan Zhu, Yanyu Zhang, Jie Xu, Wei Ren

Organizations: Department of Electrical and Computer Engineering, University of California, Riverside, CA, 92521, USA.

Abstract

3D Gaussian Splatting (3DGS) has shown promising results for 3D scene modeling using mixtures of Gaussians, yet its existing simultaneous localization and mapping (SLAM) variants typically rely on direct, deterministic pose optimization against the splat map, making them sensitive to initialization and susceptible to catastrophic forgetting as map evolves. We propose Variational Bayesian Gaussian Splatting SLAM (VBGS-SLAM), a novel framework that couples the splat map refinement and camera pose tracking in a generative probabilistic form. By leveraging conjugate properties of multivariate Gaussians and variational inference, our method admits efficient closed-form updates and explicitly maintains posterior uncertainty over both poses and scene parameters. This uncertainty-aware method mitigates drift and enhances robustness in challenging conditions, while preserving the efficiency and rendering quality of existing 3DGS. Our experiments demonstrate superior tracking performance and robustness in long sequence prediction, alongside efficient, high-quality novel view synthesis across diverse synthetic and real-world scenes.

Figures & tables

Fig. 1: VBGS-SLAM System Overview: Our method takes RGB-D images as input and initializes a point cloud via back-projection. This point cloud seeds a probabilistic 3D Gaussian map parameterized by spatial and color distribution under a generative prior model. The pipeline can be interpreted as a closed-form variational inference framework that jointly optimizes the Gaussian map and estimates the camera pose through a unified objective in Eq. ( 8 ). The online processing integrates new RGB-D image pairs, where keyframe management governs Gaussian adaption, keyframe selection, and co-visibility checking. This unified strategy enables efficient real-time SLAM by coupling mapping and tracking within a shared closed-form update loop.

Dataset	Type	Sensor	Motion Blur
Replica [ 38 ]	Synthetic	RGB-D	$\times$
TUM-RGBD [ 39 ]	Real-world	RGB-D & Acc	$\checkmark$
AR-Table [ 40 ]	Real-world	RGB-D & IMU	$\checkmark$

TABLE I: Summary of dataset characteristics. AR-TABLE dataset contains loop trajectories.

Method	Scene repr.	Input	Limitations
NICE-SLAM [ 6 ]	Implicit SDF	RGB-D	GPU - intensive
Vox-Fusion [ 11 ]	Implicit Voxel TSDF	RGB-D	Resolution
Point-SLAM [ 42 ]	Point NeRF	RGB-(D)	Sensitive to init
SplaTAM [ 13 ]	3D Gaussian splats	RGB-(D)	Opacity tuning
MonoGS [ 12 ]	3D Gaussian splats	RGB-(D)	Sensitive to init

TABLE II: Baselines comparison.

Method	r0	r1	r2	o0	o1	o2	o3	o4	Avg.
NICE-SLAM	0.97	1.31	1.07	0.88	1.00	1.06	1.10	1.13	1.07
Vox-Fusion	1.37	4.70	1.47	8.48	2.04	2.58	1.11	2.94	3.09
Point-SLAM	0.61	0.41	0.37	0.38	0.48	0.54	0.69	0.72	0.53
MnonGS	0.47	0.43	0.31	0.78	0.57	0.31	0.31	3.20	0.79
SplaTAM	0.31	0.40	0.29	0.47	0.27	0.29	0.32	0.55	0.36
Ours	0.29	0.37	0.27	0.39	0.27	0.30	0.28	0.48	0.33

TABLE III: The ATE of the estimated camera poses (cm) for five baseline methods and our VBGS-SLAM across different sequences of the Replica dataset.

Methods	fr1/	fr1/	fr1/	fr2/	fr3/	Avg.
Methods	desk	desk2	room	xyz	off.	Avg.
NICE-SLAM	4.26	4.99	34.49	31.73	3.87	15.87
Vox-Fusion	3.52	6.00	19.53	1.49	26.01	11.31
Point-SLAM	4.34	4.54	30.92	1.31	3.48	8.92
MonoGS	3.35	6.54	11.86	1.34	5.41	5.70
SplaTAM	20.21	-	-	1.47	-	-

TABLE IV: The ATE of the estimated camera poses (cm) for five baseline methods and our VBGS-SLAM across different sequences of the TUM-RGBD dataset.

Method	T1	T2	T3	T4	T5	T6	T7	Avg.
MonoGS	3.42	5.03	-	2.43	2.01	6.87	26.06	7.64
SplaTAM	18.50	6.14	3.13	-	3.22	-	-	7.75
Ours	4.56	4.22	7.57	2.61	4.85	9.11	8.70	5.94
Ours w/ IMU	4.13	3.95	6.34	2.91	4.52	8.97	9.14	5.71

TABLE V: The ATE of the estimated camera poses (cm) for five baseline methods and our VBGS-SLAM across different sequences of the AR-TABLE dataset.

Panel of Figure 7 — Fig. 2: Qualitative Comparison of Rendering image to ground truth. The top row displays results on AR-TABLE dataset, while the second row showcases the render results on TUM-RGBD dataset. Columns correspond to rendering image of ground-truth, MonoGS, SplaTAM, and VBGS-SLAM (ours), respectively.

Method	PSNR[db] $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FPS $\uparrow$
NICE-SLAM	24.42	0.809	0.233	0.54
Vox-Fusion	24.41	0.801	0.236	2.17
Point-SLAM	35.17	0.975	0.124	1.33
SplaTAM	34.11	0.97	0.10	0.115
MonoGS	37.79	0.96	0.077	0.445
Ours	37.94	0.95	0.097	0.465

TABLE VI: Quantitative View Rendering Performance on Replica Dataset

Method	PSNR[db] $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FPS $\uparrow$
NICE-SLAM	13.59	0.54	0.49	0.49
Vox-Fusion	15.54	0.63	0.50	2.17
Point-SLAM	15.63	0.66	0.53	0.22
SplaTAM	20.44	0.83	0.29	0.69
MonoGS	18.24	0.63	0.43	0.445
Ours	23.46	0.74	0.31	1.89

TABLE VII: Quantitative View Rendering Performance on TUM-RGBD Dataset

Method	PSNR[db] $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FPS $\uparrow$	GPU Uasage $\downarrow$
SplaTAM [ 13 ]	18.15	0.69	0.348	0.229	23.84
MonoGS [ 12 ]	37.79	0.96	0.077	0.445	20.86
Ours	35.24	0.78	0.079	1.374	18.57
Ours w/ IMU	36.75	0.78	0.075	1.373	18.97

TABLE VIII: Quantitative View Rendering Performance on AR-Table Dataset

3DGS SLAM

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

See all 63 papers

VarSplat: Uncertainty-aware 3D Gaussian Splatting for Robust RGB-D SLAM
Mar 10, 2026Anh Thuan Tran, Jana KoseckaSimultaneous Localization and Mapping 3DGS SLAM
Department of Computer Science, George Mason University
Geometry-Aware Online Mapping for 3D Gaussian Splatting SLAM
Aug 14, 2026Thai Luu, Quan Tran, Hieu Phan +13DGS SLAM 3D Gaussian Splatting
Cognitive Robotics Lab, Department of Electrical Engineering and Computer Science, University of Arkansas, USA. · Center for AI Research, VinUniversity, Ha Noi, Vietnam.
GS3LAM: Gaussian Semantic Splatting SLAM
Mar 29, 2026Linfei Li, Lin Zhang, Zhong Wang +13D Semantic Mapping 3DGS SLAM
School of Software Engineering, Tongji University Shanghai, China · Department of Automation, Shanghai Jiaotong University Shanghai, China

Mar 10, 2026J/K move · Enter open · S save

VarSplat: Uncertainty-aware 3D Gaussian Splatting for Robust RGB-D SLAM

Anh Thuan Tran, Jana Kosecka

Department of Computer Science, George Mason University

Simultaneous Localization and Mapping 3DGS SLAM Visual SLAM 3D Gaussian Splatting

Open paper AskPDF arXiv

Simultaneous Localization and Mapping (SLAM) with 3D Gaussian Splatting (3DGS) enables fast, differentiable rendering and high-fidelity reconstruction across diverse real-world scenes. However, existing 3DGS-SLAM approaches handle measurement reliability implicitly, making pose estimation and global alignment susceptible to drift in low-texture regions, transparent surfaces, or areas with complex reflectance properties. To this end, we introduce VarSplat, an uncertainty-aware 3DGS-SLAM system that explicitly learns per-splat appearance variance. By using the law of total variance with alpha compositing, we then render differentiable per-pixel uncertainty map via efficient, single-pass rasterization. This map guides tracking, submap registration, and loop detection toward focusing on reliable regions and contributes to more stable optimization. Experimental results on Replica (synthetic) and TUM-RGBD, ScanNet, and ScanNet++ (real-world) show that VarSplat improves robustness and achieves competitive or superior tracking, mapping, and novel view synthesis rendering compared to existing studies for dense RGB-D SLAM.

Abstract

3DGS SLAM

Explore similar work

VarSplat: Uncertainty-aware 3D Gaussian Splatting for Robust RGB-D SLAM

Geometry-Aware Online Mapping for 3D Gaussian Splatting SLAM

GS3LAM: Gaussian Semantic Splatting SLAM