Geometry-Aware Online Mapping for 3D Gaussian Splatting SLAM

Thai Luu; Quan Tran; Hieu Phan; Tuan Dang

cs.ROAug 14, 2026

Geometry-Aware Online Mapping for 3D Gaussian Splatting SLAM

Authors: Thai Luu, Quan Tran, Hieu Phan, Tuan Dang

Organizations: Cognitive Robotics Lab, Department of Electrical Engineering and Computer Science, University of Arkansas, USA. · Center for AI Research, VinUniversity, Ha Noi, Vietnam.

Abstract

Recent 3D Gaussian Splatting (3DGS) has enabled efficient photorealistic view synthesis and is rapidly being adopted in simultaneous localization and mapping (SLAM) systems for online mapping. In these systems, a Gaussian map must be expanded and refined incrementally while tracking runs in real time, so initialization and density control directly determine where limited computation and iterations are spent. This contrasts with offline 3DGS reconstruction, where such heuristics can be amortized over long optimization schedules. However, most 3DGS-SLAM pipelines inherit initialization and density-control heuristics from offline reconstruction, which can become brittle under the strict per-keyframe optimization budgets and incremental map growth of online SLAM. In this work, we revisit these heuristics in a decoupled 3DGS-SLAM setting and propose three geometry-aware methods that operate in the mapping thread: transmittance-preserving densification, camera-aware scale initialization from depth and intrinsics, and error-guided densification that focuses new primitives on high-residual regions. Our results show consistent improvements in rendering quality with negligible overhead, highlighting the coupling between photometric residuals and pose uncertainty in online SLAM. We will open-source our code to the community to foster growth and validate reproducibility.

Figures & tables

Fig. 2: System overview. Our pipeline follows Photo-SLAM’s decoupled design: ORB-SLAM3 estimates poses and selects keyframes, while an asynchronous backend optimizes a 3D Gaussian Splatting map via differentiable rasterization. mapping modules: Camera-aware Scale initialization, Transmittance-Preserving Densification, and Error-guided Densification; the tracking frontend remains unchanged.

Fig. 3: Effect of Error-guided Densification under the same online mapping budget. For each example, we show the observed keyframe (left), the baseline gradient-based densification (middle), and ours with ED (right). The residual heatmaps is brighter indicates larger error), where ED reduces holes and over-smoothed regions.

Fig. 4: Scale initialization in online SLAM. (a) k-NN scale initialization depends on sparse point density, often producing considerable splats and blurry artifacts. (b) Camera-aware Scale initializes the Gaussian size from the metric pixel footprint at the observed depth, yielding consistent, sharper primitives independent of point density.

Fig. 5: Comparison over representative views between (a) the Photo-SLAM baseline [ 10 ] and (b) our geometry-aware online mapping; PSNR (dB) is overlaid on the renderings.

Fig. 6: Additional qualitative results on TUM RGB-D. Top row: fr3/office . Bottom row: fr1/desk . Left: Ground Truth rendering. Right: Ours, demonstrating a high visual fidelity that closely matches the ground truth under the same online mapping budget.

	Scene	Baseline (Photo-SLAM)				Ours
		PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FPS $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FPS $\uparrow$
TUM RGB-D	fr1/desk	20.87	.743	.239	58.4	21.44	.761	.216	79
	fr2/xyz	22.09	.765	.169	52.9	23.27	.775	.145	72
	fr3/office	22.74	.780	.154	43.7	22.70	.774	.152	52
	Average	21.90	.763	.187	51.6	22.47	.770	.171	68
Replica	office0	38.48	.964	.050	48.6	38.41	.968	.046	58

TABLE I: Rendering quality and tracking efficiency. Evaluated on the full estimated trajectory (average of 3 runs). Our geometry-aware mapping consistently outperforms the baseline in rendering quality (PSNR, SSIM, LPIPS) across both datasets, while strictly maintaining real-time tracking performance (FPS). Best results are in bold .

Symbol	Value	Description
$\lambda$	1.0	CS coverage factor in Eq. ( 10 )
$K_{\mathrm{ED}}$	100	ED update period (mapping iterations)
$N$	$0.5\,\|\mathcal{P}_{\mathrm{err}}\|$	ED samples per update (50% of error pixels)
$\tau$	0.02	residual threshold for ED candidate pixels
$\delta$	0.5	depth-edge threshold (finite differences)
$\alpha_{0}$	0.1	initial opacity for ED-inserted Gaussians

TABLE II: Hyperparameters used by our method.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FPS $\uparrow$
BundleFusion † [ 1 ]	23.84	0.822	0.197	8.6
Nice-SLAM † [ 2 ]	26.16	0.832	0.232	2.3
ESLAM † [ 3 ]	30.59	0.866	0.162	6.7
Co-SLAM † [ 4 ]	30.25	0.864	0.175	14.6
SplaTAM † [ 12 ]	34.11	0.970	0.100	$\sim$ 0.5
GS-SLAM † [ 13 ]	34.27	0.975	0.082	$\sim$ 8.0

TABLE III: Comparison with prior systems on Replica [ 23 ] , averaged over 8 scenes. Results marked with † are from prior work, while ‡ indicates our results.

Dataset	Scene	Baseline	Ours
TUM	fr1/desk	$1.53\pm 0.01$	$1.53\pm 0.01$
	fr2/xyz	$0.31\pm 0.02$	$0.30\pm 0.01$
	fr3/office	$0.90\pm 0.02$	$0.94\pm 0.02$
	Average	$\mathbf{0.91}$	$0.92$
Replica	office0	$0.45\pm 0.02$	$0.45\pm 0.02$
	office1	$0.44\pm 0.09$	$0.37\pm 0.02$

TABLE IV: Tracking accuracy (ATE RMSE in cm, mean

\pm

std over 3 runs). TPD and CS are mapping-only; differences from Baseline lie within ORB-SLAM3 run-to-run variance.

Dataset	+TPD+CS			+TPD+CS+ED
Dataset	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
TUM RGB-D	21.91	.771	.182	21.85	.762	.184
Replica	34.72	.944	.055	34.39	.931	.055
Overall Avg	31.22	.891	.082	30.97	.874	.083

TABLE V: Rendering-quality ablation within our method (averages only), mean of 3 runs. Rendering metrics via Photo-SLAM-eval.

Iters / KF	PSNR $\uparrow$		SSIM $\uparrow$		LPIPS $\downarrow$		#Gaussians (K)
Iters / KF	Base	Ours	Base	Ours	Base	Ours	Base	Ours
50	13.70	13.80	0.575	0.578	0.598	0.588	3.50	3.50
100	12.90	14.25	0.575	0.582	0.588	0.580	1.90	1.90
200	15.90	15.70	0.613	0.618	0.498	0.490	2.30	2.40
300	16.90	16.80	0.638	0.635	0.425	0.428	2.85	2.83

TABLE VI: Iteration ablation on TUM RGB-D (average of fr1/desk and fr2/xyz). Performance of the baseline versus our proposed mapping modifications under varying per-keyframe iteration budgets.

Explore similar work

GeoGS-SLAM: Online Monocular Reconstruction Using Gaussian Splatting with Geometric Priors
Jul 13, 2026Ruilan Gao, Letian Jin, Yu ZhangSimultaneous Localization and Mapping 3D Gaussian Splatting
State Key Laboratory of Industrial Control Technology, College of Control Science and Engineering, Zhejiang University, Hangzhou, China, 310027. · Key Laboratory of Collaborative Sensing and Autonomous Unmanned Systems of Zhejiang Province, Hangzhou, China, 310027.
LightSplat: Real-Time High-Fidelity 3D Gaussian SLAM with Loop Closure
Sep 7, 2026Junze Bao, Ye Gao, Yiming Huang +5Simultaneous Localization and Mapping 3D Gaussian Splatting
School of Automation Science and Electrical Engineering, Beihang University, Beijing 100191, P. R. China, and also with Zhejiang Key Laboratory of Industrial Big Data and Robot Intelligent Systems, Hangzhou Innovation Institute, Beihang University, Hangzhou 310051, P. R. China. · School of Mechanical Engineering, Zhejiang Sci-Tech University, Hangzhou 310018, P. R. China.
GeoGS-SLAM: Geometry-Only Gaussian Splatting for Dense Monocular SLAM
Jul 8, 2026Lipu Zhou, Yaoyun Kang, Junxiang Pang +3Simultaneous Localization and Mapping 3D Gaussian Splatting
The authors are with the School of Instrument Science and Opto-electronics Engineering, Beihang University, Beijing 100191, China

Jul 13, 2026, cs.ROJ/K move · Enter open · S save

GeoGS-SLAM: Online Monocular Reconstruction Using Gaussian Splatting with Geometric Priors

Ruilan Gao, Letian Jin, Yu Zhang

State Key Laboratory of Industrial Control Technology, College of Control Science and Engineering, Zhejiang University, Hangzhou, China, 310027. · Key Laboratory of Collaborative Sensing and Autonomous Unmanned Systems of Zhejiang Province, Hangzhou, China, 310027.

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

Open paper AskPDF arXiv

SLAM methods based on 3D Gaussian Splatting (3DGS) have demonstrated impressive tracking and mapping performance, but typically require additional geometric information from external depth sensors. Meanwhile, recent SLAM systems that leverage geometric priors from pre-trained feed-forward models enable real-time dense reconstruction, yet often discard original RGB information during optimization, thus degrading overall reconstruction quality. We present GeoGS-SLAM, an online monocular dense reconstruction system that combines the 3DGS-based map representation with learned geometric priors. Given uncalibrated RGB input, we first employ a feed-forward visual geometry model to predict camera and scene priors. The Gaussian scene map is then expanded by directly sampling Gaussian primitives from both RGB input and geometric priors. Camera poses and the scene map are jointly optimized through a coarse-to-fine strategy that minimizes both photometric and geometric losses. To ensure global consistency, we further incorporate online loop closure detection and pose graph optimization. Extensive experiments across indoor and outdoor benchmarks demonstrate that GeoGS-SLAM achieves superior rendering quality and tracking accuracy compared to state-of-the-art methods while maintaining online real-time performance. Project page: https://rlgao.github.io/geogs_slam.

Abstract

Explore similar work

GeoGS-SLAM: Online Monocular Reconstruction Using Gaussian Splatting with Geometric Priors

LightSplat: Real-Time High-Fidelity 3D Gaussian SLAM with Loop Closure

GeoGS-SLAM: Geometry-Only Gaussian Splatting for Dense Monocular SLAM