LingGS: Decoupling Feed-Forward Geometry Prediction from Camera Tracking for High-Fidelity Gaussian SLAM

Da Li; Yihan Zang; Ivan Viola; Zeyu Wang; Dominik Engel

cs.CVOct 27, 2026

LingGS: Decoupling Feed-Forward Geometry Prediction from Camera Tracking for High-Fidelity Gaussian SLAM

Authors: Da Li, Yihan Zang, Ivan Viola, Zeyu Wang, Dominik Engel

Organizations: King Abdullah University of Science and Technology · The Hong Kong University of Science and Technology (Guangzhou)

Abstract

Achieving high-quality reconstruction and responsive camera tracking remains challenging for monocular RGB Gaussian splatting SLAM under a limited optimization budget. We present an online system that addresses this challenge by combining learned geometric priors with explicit geometric camera tracking, which can substantially outperform learned camera prediction in terms of tracking accuracy. A frozen streaming 3D foundation model serves as a virtual RGB-D sensor, providing depth estimation from monocular video. An explicit geometric solver then estimates camera motion through feature reprojection, and a streaming Gaussian mapper integrates the resulting observations into a dense scene representation. We further show that geometric tracking improves camera pose accuracy across representative streaming geometry backbones, demonstrating the broader applicability of our approach. Extensive experiments on indoor and outdoor datasets show that our method significantly improves reconstruction quality over existing approaches, achieving about 7 FPS on a single RTX 4090 and delivering 2-3 times the throughput of the evaluated prior SLAM systems.

Figures & tables

Figure 1: Overview of our online Gaussian splatting SLAM system. (a) A frozen streaming foundation model estimates camera intrinsics from the initial RGB frames. (b) Predicted depth and image features support geometric camera tracking through correspondence matching and reprojection error minimization. (c) The Gaussian mapper integrates RGB frames, predicted depth, and estimated poses through keyframe selection, Laplacian-guided Gaussian insertion, and 30-step refinement using RGB and depth losses with keyframe replay.

Figure 2: Ablation on refinement steps per keyframe. Means over the six Oxford Spires sequences. Rendering quality saturates at

n{\approx}75

–100 (dashed line). The filled marker is the default

n{=}30

used in all main tables.

Method	(a) 7-Scenes							(b) Oxford Spires
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	ATE $\downarrow$	$\mathrm{RPE}_{t}\downarrow$	$\mathrm{RPE}_{R}\downarrow$	FPS $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	ATE $\downarrow$	$\mathrm{RPE}_{t}\downarrow$	$\mathrm{RPE}_{R}\downarrow$	FPS $\uparrow$
MonoGS	18.94	0.681	0.501	29.90	0.776	0.290	3.28	20.40	0.635	0.571	595.47	10.58	0.321	3.21
S3PO-GS	19.76	0.713	0.470	7.30	0.480	0.213	1.74	21.63	0.699	0.506	287.81	2.802	0.134	1.31
Photo-SLAM	21.91	0.741	0.400	19.78	0.566	0.205	5.99	22.05	0.717	0.465	567.71	2.308	0.074	5.99
OTF	20.50	0.744	0.443	39.34	0.801	0.269	7.57	23.19	0.713	0.466	624.06	8.592	0.231	8.24
Ours	24.48	0.839	0.293	8.58	0.524	0.189	6.99	26.35	0.817	0.332	73.81	0.943	0.095	6.40

Table 1: Rendering quality, trajectory accuracy, and efficiency among all datasets. The best results are in bold , and the second-best results are underlined . ATE and

\mathrm{RPE}_{t}

are reported in centimeters, and

\mathrm{RPE}_{R}

is reported in degrees. For Photo-SLAM, it fails on the Pumpkin sequence of 7-Scenes and we only use 6 scenes for its average.

Figure 3: Qualitative rendering and depth comparison. One representative sequence is selected from each dataset. For 7-Scenes, TUM-D and Oxford Spires, we show both novel-view RGB rendering and rendered depth; RE10K is shown only in RGB because reference depth is unavailable. The 7-Scenes and TUM-D datasets provide dense reference depth, whereas the Oxford Spires reference depth is sparsely projected from the scanned point cloud. All methods are compared using the same valid-depth mask.

Figure 4: Qualitative trajectory comparison. We visualize one representative sequence from three datasets after trajectory alignment. Our estimated trajectories closely follow the ground truth. Because OTF is designed primarily for novel-view synthesis and estimates poses only for keyframes and held-out views, its trajectory error is computed only over this available pose subset rather than over all input frames.

Method	7-Scenes $\downarrow$	TUM-D $\downarrow$	Oxford Spires $\downarrow$
MonoGS	0.52	1.29	9.12
S3PO-GS	0.17	0.57	4.47
Photo-SLAM	0.36	1.66	8.54
OTF	1.00	2.35	14.12
Ours	0.15	0.33	3.18

Table 2: Scale-aligned rendered-depth accuracy. Mean depth L1 error on 7-Scenes, TUM-D, and Oxford Spires. All values are reported in meters; lower is better.

Pose source	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	ATE (cm) $\downarrow$	RPE t (cm) $\downarrow$	RPE R ( ∘ ) $\downarrow$	Track (ms) $\downarrow$	FPS $\uparrow$
A: Virtual stereo	26.35	0.815	0.334	79.5	1.123	0.104	16.2	6.20
B: 2D reprojection	22.90	0.709	0.446	162.1	4.723	0.140	13.9	6.51
C: ICP	21.37	0.648	0.482	307.2	15.14	0.482	138.6	5.47
D: Camera head	20.23	0.629	0.510	598.5	15.39	0.259	9.5	7.04
Ours	26.35	0.817	0.332	73.8	0.943	0.095	16.1	6.40

Table 3: Ablation on camera pose estimation over Oxford Spires. For variant C, we use Open3D’s CPU point-to-plane ICP as an unoptimized reference implementation; its runtime does not represent the computational limit of ICP.

Backbone	Pose source	ATE (m) $\downarrow$	RPE t (cm) $\downarrow$	RPE R ( ∘ ) $\downarrow$	Track (ms) $\downarrow$
XStreamVGGT ( Su et al., 2026 )	Camera head	23.277	48.95	0.4	9.6
XStreamVGGT ( Su et al., 2026 )	Geometry solver	0.773	1.44	0.085	16.1
STream3R ( Lan et al., 2026 )	Camera head	25.560	17.67	0.319	10.6
STream3R ( Lan et al., 2026 )	Geometry solver	2.332	1.74	0.098	16.1
TTT3R ( Chen et al., 2026b )	Camera head	10.57	6.52	0.204	4.2
TTT3R ( Chen et al., 2026b )	Geometry solver	7.29	2.33	0.09	16.1

Table 4: Geometry-based pose estimation across streaming backbones on Keble College-02.

3DGS SLAM

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

See all 63 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.
Flash-Mono: Feed-Forward Accelerated Gaussian Splatting Monocular SLAM
Apr 3, 2026Zicheng Zhang, Ke Wu, Xiangting Meng +33DGS SLAM 3D Gaussian Splatting
Fudan University · ShanghaiTech University · Shanghai Jiao Tong University
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

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

3DGS SLAM

Explore similar work

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

Flash-Mono: Feed-Forward Accelerated Gaussian Splatting Monocular SLAM

GaussianFlow SLAM: Monocular Gaussian Splatting SLAM Guided by GaussianFlow