RU4D-SLAM: Reweighting Uncertainty in Gaussian Splatting SLAM for 4D Scene Reconstruction

Yangfan Zhao; Hanwei Zhang; Ke Huang; Qiufeng Wang; Zhenzhou Shao; Dengyu Wu

cs.CVFeb 24, 2026

RU4D-SLAM: Reweighting Uncertainty in Gaussian Splatting SLAM for 4D Scene Reconstruction

Authors: Yangfan Zhao, Hanwei Zhang, Ke Huang, Qiufeng Wang, Zhenzhou Shao, Dengyu Wu

Organizations: Capital Normal University · Saarland University · Xi’an Jiaotong-Liverpool University · King’s College London

Abstract

Combining 3D Gaussian splatting with Simultaneous Localization and Mapping (SLAM) has gained popularity as it enables continuous 3D environment reconstruction during motion. However, existing methods struggle in dynamic environments, particularly moving objects complicate 3D reconstruction and, in turn, hinder reliable tracking. The emergence of 4D reconstruction, especially 4D Gaussian splatting, offers a promising direction for addressing these challenges, yet its potential for 4D-aware SLAM remains largely underexplored. Along this direction, we propose a robust and efficient framework, namely Reweighting Uncertainty in Gaussian Splatting SLAM (RU4D-SLAM) for 4D scene reconstruction, that introduces temporal factors into spatial 3D representation while incorporating uncertainty-aware perception of scene changes, blurred image synthesis, and dynamic scene reconstruction. We enhance dynamic scene representation by integrating motion blur rendering, and improve uncertainty-aware tracking by extending per-pixel uncertainty modeling, which is originally designed for static scenarios, to handle blurred images. Furthermore, we propose a semantic-guided reweighting mechanism for per-pixel uncertainty estimation in dynamic scenes, and introduce a learnable opacity weight to support adaptive 4D mapping. Extensive experiments on standard benchmarks demonstrate that our method substantially outperforms state-of-the-art approaches in both trajectory accuracy and 4D scene reconstruction, particularly in dynamic environments with moving objects and low-quality inputs. Code available: https://ru4d-slam.github.io

Figures & tables

Figure 2: Overview of RU4D-SLAM . RU4D-SLAM operates in three stages: pose estimation, deformation field initialization, and 4D mapping, all of which are closely linked to the uncertainty map $\boldsymbol{\beta}^{2}$ . In the pose estimation stage, $\boldsymbol{\beta}^{2}$ supports uncertainty-aware DBA tracking. Before 4D mapping, the uncertainty map is combined with SAM to form RUM, within which deformation nodes are initialized as local motion anchors by a pretrained SpaTracker model. In 4D mapping, node trajectories are propagated to Gaussians via deformation blending and optimized through IR- and AOW-guided training for joint static and dynamic rendering at each keyframe. Snowflake icons denote pre-trained, frozen modules.

Figure 3: Impact of IR on predicted uncertainty map $\boldsymbol{\beta}^{2}$ . Predicted uncertainty on a sample frame from w_x and w_r sequence in TUM dataset. Higher values are shown in red, lower values in blue, with intermediate values transitioning smoothly.

Figure 4: Comparison of dynamic object tracking across masks. Performance of YOLO-based mask [ 26 ] , flow-based mask [ 23 ] , and our RUM on three sample frame sets from w_x in TUM, pb1 in Bonn and park in Wild-SLAM.

Figure 5: Impact of AOW on rendering. Comparison of renderings without and with AOW on frames from w_x in TUM, pb1 in Bonn and park in Wild-SLAM.

Method	s_s	s_x	s_r	w_s	w_x	w_r	Avg.
w/o AOW	28.70	26.18	24.73	24.62	22.66	23.11	25.00
w/o IR	29.11	24.29	25.17	24.88	21.78	22.91	24.69
Full model	29.65	26.34	25.30	26.07	23.94	24.38	25.95

Table 1: Ablation on TUM (PSNR [dB]

\uparrow

). Performance comparison when removing AOW or IR individually. Best per column is bold , second best is underlined .

Method	Type	s_s	s_x	s_r	w_s	w_x	w_r	Avg.
w/o IR	Static	378	672	213	347	525	597	459
w/o IR	Dynamic	323	213	514	441	772	893	526
Full model	Static	159	274	368	160	263	360	264
Full model	Dynamic	100	75	141	112	128	298	142

Table 2: Number of dynamic and static Gaussians after training on TUM (in thousands).

Method	Metric	s_s	s_x	s_r	w_s	w_x	w_r	Avg.
MonoGS [ 32 ]	PSNR $\uparrow$	19.95	23.92	16.99	16.47	14.02	15.12	17.74
	SSIM $\uparrow$	0.739	0.803	0.572	0.604	0.436	0.497	0.608
	LPIPS $\downarrow$	0.213	0.182	0.405	0.355	0.581	0.560	0.382
Gaussian-SLAM [ 64 ]	PSNR $\uparrow$	18.57	19.22	16.75	14.91	14.67	14.50	16.43
	SSIM $\uparrow$	0.848	0.796	0.652	0.607	0.483	0.467	0.642
	LPIPS $\downarrow$	0.291	0.326	0.521	0.489	0.626	0.630	0.480

Table 3: Rendering quality on TUM. (PSNR [dB]

\uparrow

, SSIM

\uparrow

, and LPIPS

\downarrow

)

\uparrow

indicates higher is better;

\downarrow

indicates lower is better. Best results are bold , second best are underlined .

Method	Metric	bal1	bal2	ps1	ps2	sy1	sy2	pb1	pb2	pb3	Avg.
MonoGS [ 32 ]	PSNR $\uparrow$	21.35	20.22	20.53	20.09	22.03	20.55	20.76	19.38	24.81	21.06
	SSIM $\uparrow$	0.803	0.758	0.779	0.718	0.766	0.841	0.748	0.753	0.857	0.780
	LPIPS $\downarrow$	0.316	0.354	0.408	0.426	0.328	0.521	0.428	0.372	0.243	0.342
Gaussian-SLAM [ 64 ]	PSNR $\uparrow$	20.45	18.55	19.60	19.09	21.04	21.35	19.99	20.35	21.22	20.18
	SSIM $\uparrow$	0.792	0.718	0.744	0.719	0.784	0.837	0.750	0.768	0.814	0.769
	LPIPS $\downarrow$	0.457	0.480	0.484	0.496	0.402	0.364	0.509	0.493	0.441	0.458

Table 4: Rendering quality on Bonn. (PSNR [dB]

\uparrow

, SSIM

\uparrow

, and LPIPS

\downarrow

)

\uparrow

indicates higher is better;

\downarrow

indicates lower is better. Best results are bold , second best are underlined .

Method	Metric	park	piano	shop	street	tow	wall	wand	Avg.
MonoGS [ 32 ]	PSNR $\uparrow$	21.64	14.82	14.42	17.86	16.38	14.66	13.48	16.18
	SSIM $\uparrow$	0.777	0.353	0.326	0.601	0.667	0.361	0.302	0.484
	LPIPS $\downarrow$	0.317	0.473	0.578	0.350	0.491	0.638	0.533	0.483
4DGS SLAM [ 26 ]	PSNR $\uparrow$	21.96	21.13	19.57	19.63	22.62	16.55	18.62	20.01
	SSIM $\uparrow$	0.797	0.705	0.608	0.708	0.806	0.432	0.560	0.659
	LPIPS $\downarrow$	0.306	0.269	0.380	0.315	0.249	0.688	0.387	0.371

Table 5: Rendering quality on Wild-SLAM. (PSNR [dB]

\uparrow

, SSIM

\uparrow

, and LPIPS

\downarrow

)

\uparrow

indicates higher is better;

\downarrow

indicates lower is better. Best results are bold , second best are underlined .

Method	s_s	s_x	s_r	w_s	w_x	w_r	Avg.
MonoGS [ 32 ]	0.48	1.70	6.10	21.90	30.70	34.20	15.80
Gaussian-SLAM [ 64 ]	0.72	1.40	21.02	91.50	168.10	152.00	72.40
SplaTAM [ 16 ]	0.52	1.50	11.80	83.20	134.20	142.30	62.20
RoDyn-SLAM [ 13 ]	1.50	5.60	5.70	1.70	8.30	8.10	5.10
DG-SLAM [ 59 ]	0.72	1.00	3.49	0.60	1.60	4.30	1.95
WildGS-SLAM* [ 66 ]	0.51	1.70	2.19	0.49	2.19	3.11	1.70

Table 6: Pose accuracy on TUM (ATE [cm]

\downarrow

\downarrow

indicates lower is better. Best values per column are bold , second best are underlined . WildGS-SLAM* is re-evaluated without final bundle adjustment to avoid pose errors from dynamic objects.

Method	bal1	bal2	ps1	ps2	sy1	sy2	pb1	pb2	pb3	Avg.
MonoGS [ 32 ]	29.60	22.10	54.50	36.90	68.50	0.56	71.50	10.70	3.60	33.10
Gaussian-SLAM [ 64 ]	66.90	32.80	107.20	114.40	111.80	164.80	69.90	53.80	37.90	84.30
SplaTAM [ 16 ]	32.90	30.40	77.80	116.70	59.50	66.70	91.90	18.50	17.10	56.80
RoDyn-SLAM [ 13 ]	7.90	11.50	14.50	13.80	1.30	1.40	4.90	6.20	10.20	7.90
DG-SLAM [ 59 ]	3.70	4.10	4.50	6.90	1.12	7.71	24.22	4.59	4.67	6.83
WildGS-SLAM* [ 66 ]	3.09	2.80	4.35	5.39	0.84	0.55	2.07	3.16	2.81	2.78

Table 7: Pose accuracy on Bonn (ATE [cm]

\downarrow

\downarrow

indicates lower is better. Best values per column are bold , second best are underlined . WildGS-SLAM* is re-evaluated without final bundle adjustment to avoid pose errors from dynamic objects.

3DGS SLAM

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

See all 63 papers

SCOUT-SLAM: Structurally-Coupled Dual Uncertainty-Aware 3DGS SLAM in the Wild
Sep 13, 2026Kumaran Karthik, Pramat Shastri Jois, Suresh Sundaram3DGS SLAM Simultaneous Localization and Mapping
Indian Institute of Science
Flow4DGS-SLAM: Optical Flow-Guided 4D Gaussian Splatting SLAM
Apr 24, 2026Yunsong Wang, Gim Hee LeeSimultaneous Localization and Mapping 3D Gaussian Splatting
School of Computing, National University of Singapore
RoSe-SLAM: Robust Semantic-Aware Gaussian Splatting SLAM from Dynamic Monocular Videos
Aug 29, 2026Wenting Wang, Jiaxin Guo, Wenzhen Dong +3Dynamic SLAM Visual SLAM
Department of Mechanical and Automation Engineering, The Chinese University of Hong Kong, Hong Kong SAR. · Department of Mechanical, Aerospace and Civil Engineering, The University of Manchester, Manchester, UK. · Centre for Perceptual and Interactive Intelligence (CPII) Limited, Hong Kong SAR.

Sep 13, 2026J/K move · Enter open · S save

SCOUT-SLAM: Structurally-Coupled Dual Uncertainty-Aware 3DGS SLAM in the Wild

Kumaran Karthik, Pramat Shastri Jois, Suresh Sundaram

Indian Institute of Science

3DGS SLAM Simultaneous Localization and Mapping 3D Reconstruction Camera Pose Estimation

Open paper AskPDF arXiv

Recently, 3D Gaussian Splatting SLAM (3DGS-SLAM) has gained significant momentum in simultaneous localization and 3DGS scene reconstruction. In real-world scenarios with rapid camera motion and cluttered dynamic environments, existing methods rely on the stability of the underlying scene reconstruction to model uncertainty. This leads to a circular dependency between camera tracking accuracy and reconstruction quality: reconstruction instabilities degrade uncertainty modeling, which affects accurate camera tracking and static scene reconstruction. To address this, the paper proposes SCOUT-SLAM, a structurally-coupled dual-uncertainty framework in which both uncertainties are estimated from a shared base network. A low-rank adaptation of this network, trained on multi-view feature consistency, estimates a tracking uncertainty that does not depend solely on the reconstruction quality. A spatially-adaptive prior modulates the network's training objective so that reconstruction instability does not inflate uncertainty on static regions, keeping the shared representation intact for both branches. Evaluations on dynamic benchmarks (TUM RGB-D, Bonn Dynamic, Wild-SLAM MoCap) demonstrate that SCOUT-SLAM achieves state-of-the-art camera tracking accuracy and artifact-free static scene reconstruction.

Abstract

3DGS SLAM

Explore similar work

SCOUT-SLAM: Structurally-Coupled Dual Uncertainty-Aware 3DGS SLAM in the Wild

Flow4DGS-SLAM: Optical Flow-Guided 4D Gaussian Splatting SLAM

RoSe-SLAM: Robust Semantic-Aware Gaussian Splatting SLAM from Dynamic Monocular Videos