UniGS-SLAM: Monocular Gaussian SLAM with Enhanced Geometry and Open-Vocabulary Semantics

Xingyuan Ding; Sicheng Yu; Kaiyong Zhao; Hao Wang

cs.CVNov 1, 2026

UniGS-SLAM: Monocular Gaussian SLAM with Enhanced Geometry and Open-Vocabulary Semantics

Authors: Xingyuan Ding, Sicheng Yu, Kaiyong Zhao, Hao Wang

Organizations: Shanghai Jiao Tong University · The Hong Kong University of Science and Technology (Guangzhou) · XGRIDS

Abstract

Monocular semantic SLAM aims to reconstruct scene geometry and build semantic maps from RGB observations. Existing Gaussian-based methods often attach semantic labels or features to evolving Gaussian primitives, making semantic fusion difficult to maintain and increasing reconstruction overhead. Hybrid Gaussian--voxel representations provide more stable semantic support, but often rely on RGB-D input or known poses and do not revise historical semantic evidence after geometry updates. We present UniGS-SLAM, an RGB-only framework that decouples open-vocabulary semantics from Gaussian reconstruction while preserving bidirectional interaction. Foundation priors enhance monocular depth and pose estimation, while semantic observations are fused into an independent sparse voxel map with observation provenance, enabling historical contributions to be revised as reconstruction evolves. The fused semantics further guide Gaussian filtering and local surface refinement. Experiments on Replica, ScanNet, and KITTI-360 show leading tracking, rendering, and geometric reconstruction performance among representative RGB-only methods, together with strong open-set semantic mapping on Replica and ScanNet. We also achieve a 2.37-fold reconstruction speedup over the fastest compared semantic SLAM baseline. Project page: https://3dagentworld.github.io/UniGS-SLAM/.

Figures & tables

Figure 1: Our RGB-only framework reconstructs detailed geometry and open-vocabulary semantics (left), achieves strong quality and efficiency (middle), and supports 3D scene understanding (right).

Figure 2: Overview of UniGS-SLAM. Foundation geometry priors enhance monocular reconstruction. An independent voxel map fuses semantic observations and revises historical evidence after geometry updates. Semantic feedback guides Gaussian floater filtering and surface refinement.

Figure 3: RGB rendering comparisons on ScanNet and KITTI-360. Our method recovers finer scene details and achieves higher rendering quality.

Panel of Figure 4 — Figure 4: Mesh comparisons on Replica and ScanNet, showing cleaner surfaces and finer details.

Panel of Figure 5 — Figure 5: Semantic maps on Replica, showing coherent regions consistent with ground truth.

Method [cm] $\downarrow$	PSNR [dB] $\uparrow$	L1 depth [cm] $\downarrow$	ATE RMSE [cm] $\downarrow$
Foundation Prior (DA3)	33.68	1.964	2.010
+ Context & Anchor Alignment	34.79	1.310	0.713
+ Depth Fusion	40.34	1.083	0.623
+ Long-Horizon Guidance	42.61	1.023	0.207

Table 6: Geometry enhancement ablation.

Method	mAcc $\uparrow$	mIoU $\uparrow$
w/o Historical Revision	55.85	45.43
Reprojection w/o Retraction	56.90	46.25
Ours	60.80	49.86

Table 7: Semantic revision ablation.

Method	Acc. [cm] $\downarrow$	F-score [%] $\uparrow$
w/o both	13.76	48.86
+ Floater Filtering	10.08	50.28
+ Surface Refinement	9.79	50.97

Table 8: Semantic-guided refinement.

Method	Time (min) $\downarrow$
SNI-SLAM	48.32
SGS-SLAM	118.70
SemGauss-SLAM	132.79
GS3LAM	172.40
Open-set SLAM	90.10
Ours	20.40

Table 9: Reconstruction time.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Scene understanding applications.

Sequence	Frame range	Frames
0002	13419–13718	300
0004	1946–2245	300
0004	6192–6491	300
0005	2030–2329	300

Appendix

Table 10: KITTI-360 segments with inclusive source-frame ranges.

Method	PSNR [dB] $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	L1 Depth [cm] $\downarrow$
w/o both	36.79	0.958	0.036	1.564
+ Geometry Enhancement	38.37	0.974	0.030	1.260
+ Semantic-Guided Refinement	38.65	0.975	0.026	1.170

Appendix

Table 11: Progressive ablation of geometry enhancement and semantic-guided refinement, averaged over all eight Replica scenes.

Refinement Variant	Acc. [cm] $\downarrow$	F-score [%] $\uparrow$
No Refinement	13.76	48.86
Geometry-Only Refinement	12.07	49.14
Semantic-Guided Refinement (ours)	9.79	50.97

Appendix

Table 12: Ablation of semantic guidance in Gaussian geometry refinement on ScanNet scene0207. The geometry-only variant retains the same geometric refinement while removing semantic guidance.

Method	Metric	ro-0	ro-1	ro-2	of-0	of-1	of-2	of-3	of-4	Avg.
Splat-SLAM	PSNR $\uparrow$	30.75	32.69	32.63	39.00	36.15	33.69	33.52	34.73	34.15
	SSIM $\uparrow$	0.914	0.930	0.944	0.973	0.964	0.930	0.943	0.959	0.945
	LPIPS $\downarrow$	0.091	0.076	0.070	0.046	0.066	0.113	0.066	0.066	0.074
ARTDECO	PSNR $\uparrow$	29.28	28.34	30.44	35.63	34.38	27.88	28.29	28.40	30.33
	SSIM $\uparrow$	0.883	0.875	0.920	0.951	0.936	0.915	0.901	0.902	0.910
	LPIPS $\downarrow$	0.172	0.182	0.153	0.137	0.189	0.188	0.173	0.174	0.171

Appendix

Table 13: Rendering quality on Replica.

Method	ro-0	ro-1	ro-2	of-0	of-1	of-2	of-3	of-4	Avg.
Splat-SLAM	2.93	2.17	2.19	2.50	1.79	2.83	2.76	2.57	2.468
ARTDECO	7.11	6.70	6.04	4.37	4.42	10.62	8.83	52.43	12.565
S3PO-GS	9.84	12.02	7.94	5.62	20.92	10.99	12.66	12.21	11.523
DROID-Splat	4.58	2.75	3.78	2.06	2.01	10.82	15.60	3.93	5.690
FreeOcc	16.90	8.24	14.22	7.90	6.37	13.05	19.33	16.56	12.820
HI-SLAM2	1.93	1.17	1.47	1.32	0.86	1.54	2.19	1.75	1.528

Appendix

Table 14: Depth reconstruction error on Replica (L1 depth

\downarrow

[cm]).

Figure 7: Additional qualitative comparisons of reconstructed meshes against baseline methods and the ground truth.

Method	Metric	0000	0059	0106	0169	0181	0207	Avg.
Splat-SLAM	PSNR $\uparrow$	23.29	21.30	22.49	24.66	17.93	22.72	22.07
	SSIM $\uparrow$	0.800	0.793	0.803	0.801	0.762	0.794	0.792
	LPIPS $\downarrow$	0.220	0.216	0.248	0.218	0.249	0.230	0.230
ARTDECO	PSNR $\uparrow$	19.37	19.99	19.36	20.48	19.02	20.16	19.73
	SSIM $\uparrow$	0.732	0.794	0.812	0.806	0.806	0.752	0.784
	LPIPS $\downarrow$	0.363	0.343	0.341	0.324	0.397	0.372	0.357

Appendix

Table 15: Rendering quality on ScanNet.

Method	0000	0059	0106	0169	0181	0207	Avg.
Splat-SLAM	8.89	12.57	24.53	12.36	16.83	12.80	14.66
ARTDECO	8.45	15.67	18.72	15.33	12.71	15.92	14.47
S3PO-GS	43.78	20.05	22.70	83.01	50.43	34.59	42.43
DROID-Splat	17.59	17.33	19.00	31.13	12.65	18.70	19.40
FreeOcc	12.66	26.35	31.87	19.87	11.23	16.31	19.72
HI-SLAM2	5.24	10.16	14.92	13.83	6.90	11.75	10.47

Appendix

Table 16: Depth reconstruction error on ScanNet (L1 depth

\downarrow

[cm]).

Method	Metric	scene01	scene02	scene03	scene04	Avg.
Splat-SLAM	PSNR $\uparrow$	22.82	24.89	22.77	23.13	23.40
	SSIM $\uparrow$	0.740	0.744	0.683	0.748	0.729
	LPIPS $\downarrow$	0.331	0.353	0.430	0.353	0.367
ARTDECO	PSNR $\uparrow$	20.74	22.81	23.03	21.43	22.00
	SSIM $\uparrow$	0.698	0.705	0.720	0.731	0.714
	LPIPS $\downarrow$	0.268	0.284	0.276	0.285	0.278

Appendix

Table 17: Rendering quality on KITTI-360.

Figure 8: Additional semantic reconstruction results. From left to right: RGB images, depth maps, closed-set Gaussian semantics, open-set voxel semantics, and ground-truth semantic annotations.

Figure 9: Additional RGB rendering comparisons on ScanNet and KITTI-360. Enlarged regions highlight local appearance details.

Visual SLAM

Visual simultaneous localization and mapping (SLAM) estimates camera motion while building maps from images for navigation and scene understanding.

See all 112 papers

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.
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
VCS-SLAM: Visibility-Consistent Semantic Fusion for 3D Gaussian SLAM
Jun 28, 2026Raman Jha, Shuaihang Yuan, Yi Fang3D Gaussian Splatting 3DGS SLAM
New York University, Electrical & Computer Engineering Dept., Brooklyn, NY 11201, USA. · Embodied AI and Robotics (AIR) Lab, NYU Abu Dhabi, UAE. · New York University Abu Dhabi, Electrical Engineering, Abu Dhabi 129188, UAE.

Jul 13, 2026J/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

Visual SLAM

Explore similar work

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

GS3LAM: Gaussian Semantic Splatting SLAM

VCS-SLAM: Visibility-Consistent Semantic Fusion for 3D Gaussian SLAM