MonoEM-GS: Monocular Expectation-Maximization Gaussian Splatting SLAM

Evgenii Kruzhkov; Sven Behnke

cs.ROApr 12, 2026

MonoEM-GS: Monocular Expectation-Maximization Gaussian Splatting SLAM

Authors: Evgenii Kruzhkov, Sven Behnke

Organizations: 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

Abstract

Feed-forward geometric foundation models can infer dense point clouds and camera motion directly from RGB streams, providing priors for monocular SLAM. However, their predictions are often view-dependent and noisy: geometry can vary across viewpoints and under image transformations, and local metric properties may drift between frames. We present MonoEM-GS, a monocular mapping pipeline that integrates such geometric predictions into a global Gaussian Splatting representation while explicitly addressing these inconsistencies. MonoEM-GS couples Gaussian Splatting with an Expectation--Maximization formulation to stabilize geometry, and employs ICP-based alignment for monocular pose estimation. Beyond geometry, MonoEM-GS parameterizes Gaussians with multi-modal features, enabling in-place open-set segmentation and other downstream queries directly on the reconstructed map. We evaluate MonoEM-GS on 7-Scenes, TUM RGB-D and Replica, and compare against recent baselines.

Figures & tables

Fig. 1: MonoEM-GS constructs a map in which each Gaussian aggregates measurements over time. Despite contradictory point-cloud predictions at timestamps $t$ and $t{+}1$ , the Gaussians remain consistent, with only a single update step needed to incorporate new observations.

Fig. 2: Single iteration of processing a new input image by MonoEM-GS. For each new RGB input, MapAnything [ 27 ] produces a dense point cloud prediction $o^{N}_{t}$ . We align it with the new prediction for frame N-1: $o^{N-1}_{t}$ using ICP [ 15 , 16 ] . Since $o^{N-1}_{t}$ and $o^{N-1}_{t-1}$ are created from the same frame at different timestamps, we employ the found correspondences between $o^{N-1}_{t}$ and $o^{N}_{t}$ to register $o^{N}_{t}$ to $o^{N-1}_{t-1}$ . After the initial registration has been found, we perform colored ICP [ 14 ] alignment of $o^{N}_{t}$ to the map and estimate the camera pose with respect to the map. The aligned prediction $o^{N}_{t}$ is fused into the map and appended to the FIFO buffer.

Fig. 3: Views of the current prediction alignment with the map. View 1: misaligned current prediction $o^{N}_{t}$ ; View 2: predictions $o^{N-1}_{t}$ and $o^{N-1}_{t-1}$ for the same buffered image at different timestamps; View 3: the found associations between $o^{N}_{t}$ and $o^{N-1}_{t-1}$ ; View 4: current prediction $o^{N}_{t}$ aligned to the map.

Fig. 4: Multi-domain scene reconstructions produced by MonoEM-GS. A : Gaussian Splatting created by the proposed approach; B : Mesh reconstructed from Gaussian centers and normals; C : DINOv3 [ 29 ] feature map; D : Open-set 3D semantic map.

Parameter	Value	Parameter	Value
Buffer size $\lvert\mathcal{S}\rvert$	10	$\lambda$	128
Normal distance $\tau_{n}$	0.5	$\kappa_{1}$	5
$\tau_{\Sigma}$	4.6	$\kappa_{2}$	5
Gaussians per point $K$	16	$\tau_{\pi}$ (in $\log$ -space)	8
Gaussians per point $K_{2\textrm{D}}$	3	$\tau_{\Sigma 2D}$	4.6

TABLE I: MonoEM-GS Parameters.

	7-Scenes [ 1 ]						TUM RGB-D [ 2 ]
	Meters [m]			Percentages [%]			Meters [m]			Percentages [%]
Methods	Acc. ${}\mkern 2.0mu\downarrow$	Comp. ${}\mkern 2.0mu\downarrow$	Trajectory / ATE RMSE ${}\mkern 2.0mu\downarrow$	F1 ${}\mkern 2.0mu\uparrow$	Normals ${}\mkern 2.0mu\uparrow$	Scale ${}\mkern 2.0mu\uparrow$	Acc. ${}\mkern 2.0mu\downarrow$	Comp. ${}\mkern 2.0mu\downarrow$	Trajectory / ATE RMSE ${}\mkern 2.0mu\downarrow$	F1 ${}\mkern 2.0mu\uparrow$	Normals ${}\mkern 2.0mu\uparrow$	Scale ${}\mkern 2.0mu\uparrow$
VGGT-SLAM [ 11 ]	0.06	0.27	0.07	70.7	34.0	47.2	0.15	0.27	0.05	63.2	46.2	74.8
VGGT-SLAM2 [ 12 ]	0.09	0.11	0.07	92.8	55.5	57.7	0.11	0.21	0.04	79.6	62.2	65.8
MASt3R-SLAM [ 13 ]	0.11	0.16	0.05	85.0	41.7	88.4	0.08	0.19	0.03	83.8	44.2	86.1
MapAnything [ 27 ]	0.13	0.14	0.1	84.7	35.4	71.9	0.15	0.24	0.13	65.9	38.9	75.4

TABLE II: Mapping and Localization Evaluation on 7-Scenes [ 1 ] and TUM RGB-D [ 2 ] Datasets.

Fig. 5: Qualitative comparison of VGGT-SLAM2 [ 12 ] ( left ) and MonoEM-GS ( right ) maps. A – A : VGGT-SLAM2 [ 12 ] point cloud and MonoEM-GS Gaussian centers on the TUM RGB-D [ 2 ] “desk2” sequence; B – B : VGGT-SLAM2 [ 12 ] point cloud and MonoEM-GS Gaussian centers on the 7-Scenes [ 1 ] “office” sequence; C – C : meshes reconstructed for both methods from the point clouds in B and B , respectively.

Fig. 6: Reconstructed Gaussians (upper row) and corresponding meshes (bottom row) created from the Gaussian centers and normals.

Fig. 7: MonoEM-GS Gaussian centers colored by their predicted classes ( right ) and the nearest ground-truth points colored by their GT classes ( left ).

Methods	mIoU [%]	f-mIoU [%]	Acc [%]	Trajectory / ATE RMSE [cm]
				Trajectory / ATE RMSE [cm]
ConceptGraphs [ 37 ]	16.7	35.75	33.7	N/A
RayFronts [ 32 ]	27.7	43.47	54.5	N/A
OVO/ORB-SLAM2 [ 31 ]	27.1	60.2	39.1	1.9
OML [ 30 ]	32.1	49.6	56.2	35
MonoEM-GS (ours)	31.5	63.2	47.4	13.1

TABLE III: Replica 3D [ 3 ] semantic segmentation evaluation.

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
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
SG^2SLAM: Spatial Geometry-Informed 3D Gaussian Splatting SLAM for Monocular Outdoor Navigation
Oct 26, 2026Ziang Guo, Didar Seyidov, Yixiao Zhou +53DGS SLAM Visual SLAM
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. +2

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

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

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