SGAD-SLAM: Splatting Gaussians at Adjusted Depth for Better Radiance Fields in RGBD SLAM

Pengchong Hu; Zhizhong Han

cs.CVMar 22, 2026

SGAD-SLAM: Splatting Gaussians at Adjusted Depth for Better Radiance Fields in RGBD SLAM

Authors: Pengchong Hu, Zhizhong Han

Organizations: Machine Perception Lab, Wayne State University, Detroit, USA

Abstract

3D Gaussian Splatting (3DGS) has made remarkable progress in RGBD SLAM. Current methods usually use 3D Gaussians or view-tied 3D Gaussians to represent radiance fields in tracking and mapping. However, these Gaussians are either too flexible or too limited in movements, resulting in slow convergence or limited rendering quality. To resolve this issue, we adopt pixel-aligned Gaussians but allow each Gaussian to adjust its position along its ray to maximize the rendering quality, even if Gaussians are simplified to improve system scalability. To speed up the tracking, we model the depth distribution around each pixel as a Gaussian distribution, and then use these distributions to align each frame to the 3D scene quickly. We report our evaluations on widely used benchmarks, justify our designs, and show advantages over the latest methods in view rendering, camera tracking, runtime, and storage complexity. Please see our project page for code and videos at https://machineperceptionlab.github.io/SGAD-SLAM-Project .

Figures & tables

Figure 1: Overview of our method. Our tracking strategy, illustrated in (a)-(e), employs geometry Gaussians to represent the scene structure, achieving robustness and efficiency. During mapping, we learn appearance Gaussians for rendering, where the depth offset at each pixel is progressively adjusted for better rendering, improving scalability and efficiency. We visualize the Gaussian centers for better readability.

Figure 2: Illustration of Offsets which place Gaussians either in front (blue) or behind (red) the depth points.

Figure 3: Advantages of pixel-aligned Gaussians at adjusted depth over others in rendering comparison.

Panel of Figure 4 — Figure 4: (a) Pixel-aligned Gaussians at adjusted depth. (b) Illustration of Scale Normalization. Each Gaussian is shown with a minimum probability of 0.99.

Table 5

Figure 5: Error map comparisons in rendering on Replica [ 66 ] .

Figure 6: Reconstruction comparisons on TUM-RGBD [ 67 ] .

Neural Implicit Fields
Method	Rm0	Rm1	Rm2	Off0	Off1	Off2	Off3	Off4	Avg.
NICE-SLAM [ 103 ]	1.69	2.04	1.55	0.99	0.90	1.39	3.97	3.08	1.95
DF-Prior [ 26 ]	1.39	1.55	2.60	1.09	1.23	1.61	3.61	1.42	1.81
Vox-Fusion [ 82 ]	0.27	1.33	0.47	0.70	1.11	0.46	0.26	0.58	0.65
ESLAM [ 33 ]	0.71	0.70	0.52	0.57	0.55	0.58	0.72	0.63	0.63
Point-SLAM [ 59 ]	0.61	0.41	0.37	0.38	0.48	0.54	0.72	0.63	0.52

Table 3: Tracking results in ATE RMSE

\downarrow[\mathrm{cm}]

on Replica.

Neural Implicit Fields
Method	fr1/desk	fr2/xyz	fr3/office	Avg.
NICE-SLAM [ 103 ]	4.3	31.7	3.9	13.3
Vox-Fusion [ 82 ]	3.5	1.5	26.0	10.3
Point-SLAM [ 59 ]	4.3	1.3	3.5	3.0
Loopy-SLAM $*$ [ 41 ]	3.8	1.6	3.4	2.9
3D Gaussian Splatting

Table 4: Tracking results in ATE RMSE

\downarrow[\mathrm{cm}]

on TUM-RGBD.

Figure 8: Rendering comparisons on ScanNet [ 14 ] .

Table 12

Method	SplaTAM [ 34 ]	Gaussian- SLAM [ 87 ]	VTGS- SLAM [ 27 ]	Loop Splat $*$ [ 102 ]	Ours
Training views	26.71	30.38	32.22	30.20	36.73
Novel views	17.82	21.27	21.46	21.30	22.32

Table 7: Rendering comparisons in PSNR

\uparrow

on ScanNet++ [ 83 ] .

	Tracking /Frame(s) $\downarrow$	Mapping /Frame(s) $\downarrow$	Total /Frame(s) $\downarrow$	NUM1	NUM2
NICE-SLAM [ 103 ]	1.06	1.15	2.21	-	-
Point-SLAM [ 59 ]	1.11	3.52	4.63	-	-
SplaTAM [ 34 ]	2.70	4.89	7.59	5832 $K$	5832 $K$
Gaussian-SLAM [ 87 ]	0.83	0.93	1.76	32592 $K$	1983 $K$
GS-ICP SLAM [ 20 ]	0.03	1.02	1.05	1544 $K$	1544 $K$
Ours	0.01	0.89	0.90	326400 $K$	816 $K$

Table 8: Runtime and Memory Usage on Replica. Ours † : Parallel on 8 GPUs. NUM1: Total Number of Gaussians. NUM2: Max Number of Learnable Gaussians.

	Mode1	Mode2	Mode3	Ours
Num of Gaussians	4450 $K$	205 $K$	205 $K$	205 $K$
Num of attributes	14	14	5	6
PSNR $\uparrow$	20.90	37.67	37.16	38.60
SSIM $\uparrow$	0.771	0.995	0.995	0.997
LPIPS $\downarrow$	0.216	0.017	0.019	0.012

Table 9: Ablation study on Gaussian modeling on TUM-RGBD [ 67 ] . Mode1: All Gaussians Movable in the Whole Scene; Mode2: Gaussians Movable in the View Frustum; Mode3: Gaussians Fixed at Depth; Ours: Gaussians Movable Along the Ray.

Figure 9: Training view rendering comparisons on ScanNet++.

Figure 10: Novel view rendering comparisons on ScanNet++.

	10 $\%$	20 $\%$	30 $\%$	40 $\%$	Ours
PSNR $\uparrow$	39.86	39.86	39.86	39.86	39.86
SSIM $\uparrow$	0.997	0.997	0.997	0.997	0.997
LPIPS $\downarrow$	0.013	0.013	0.013	0.013	0.012
ATE RMSE $\downarrow$	2.1	2.1	2.2	2.2	2.0

Table 10: Ablation study on robustness to additional noise in rendering and tracking on fr3/office in TUM-RGBD [ 67 ] .

	Point2point w/o scale(GICP)	Point2point w/ Ellipse scale	Point2surf w/o scale	Point2surf w/ Plane scale	Ours
ATE RMSE $\downarrow$	106.1	17.0	104.6	27.0	2.0

Table 11: Ablation study on metrics for correspondence and scale regularization on TUM-RGBD. Ours: Point2surf & Ellipse scale.

Figure 11: Visual comparisons in rendering on TUM-RGBD [ 67 ] . Please watch our video for more comparisons in rendering.

Figure 12: Visual comparisons in rendering on ScanNet [ 14 ] . Please watch our video for more comparisons in rendering.

Figure 13: Visual comparisons in reconstruction on Replica [ 66 ] .

Figure 14: Visual comparisons in reconstruction and camera tracking on Replica [ 66 ] . Please watch our video for more comprehensive results.

Figure 15: Visual comparisons in reconstruction on TUM-RGBD [ 67 ] .

Figure 16: Visual comparisons in reconstruction on ScanNet [ 14 ] .

Figure 17: Visual comparisons in training views rendering on ScanNet++ [ 83 ] . Please watch our video for more comparisons in rendering.

Neural Implicit Fields
Method	Metric	Rm0	Rm1	Rm2	Off0	Off1	Off2	Off3	Off4	Avg.
NICE-SLAM [ 103 ]	PSNR $\uparrow$	22.12	22.47	24.52	29.07	30.34	19.66	22.23	24.94	24.42
	SSIM $\uparrow$	0.689	0.757	0.814	0.874	0.886	0.797	0.801	0.856	0.809
	LPIPS $\downarrow$	0.330	0.271	0.208	0.229	0.181	0.235	0.209	0.198	0.233
Vox-Fusion [ 82 ]	PSNR $\uparrow$	22.39	22.36	23.92	27.79	29.83	20.33	23.47	25.21	24.41
	SSIM $\uparrow$	0.683	0.751	0.798	0.857	0.876	0.794	0.803	0.847	0.801

Table 12: Rendering performance comparisons in PSNR

\uparrow

, SSIM

\uparrow

, and LPIPS

\downarrow

on Replica [ 66 ] .

*

indicates methods relying on pre-trained data-driven priors.

Neural Implicit Fields
Method	Metric	Rm0	Rm1	Rm2	Off0	Off1	Off2	Off3	Off4	Avg.
NICE-SLAM [ 103 ]	Depth L1 [cm] $\downarrow$	1.81	1.44	2.04	1.39	1.76	8.33	4.99	2.01	2.97
NICE-SLAM [ 103 ]	F1 $[\mathrm{\%}]\uparrow$	45.0	44.8	43.6	50.0	51.9	39.2	39.9	36.5	43.9
Vox-Fusion [ 82 ]	Depth L1 [cm] $\downarrow$	1.09	1.90	2.21	2.32	3.40	4.19	2.96	1.61	2.46
Vox-Fusion [ 82 ]	F1 $[\mathrm{\%}]\uparrow$	69.9	34.4	59.7	46.5	40.8	51.0	64.6	50.7	52.2
ESLAM [ 33 ]	Depth L1 [cm] $\downarrow$	0.97	1.07	1.28	0.86	1.26	1.71	1.43	1.06	1.18

Table 13: Reconstruction performance comparison in Depth L1 [cm]

\downarrow

and F1

[\mathrm{\%}]\uparrow

on Replica [ 66 ] .

*

indicates methods relying on pre-trained data-driven priors.

Neural Implicit Fields
Method	Metric	fr1/desk	fr2/xyz	fr3/office	Avg.
NICE-SLAM [ 103 ]	PSNR $\uparrow$	13.83	17.87	12.89	14.86
	SSIM $\uparrow$	0.569	0.718	0.554	0.614
	LPIPS $\downarrow$	0.482	0.344	0.498	0.441
Vox-Fusion [ 82 ]	PSNR $\uparrow$	15.79	16.32	17.27	16.46
	SSIM $\uparrow$	0.647	0.706	0.677	0.677

Table 14: Rendering performance comparison in PSNR

\uparrow

, SSIM

\uparrow

, and LPIPS

\downarrow

on TUM-RGBD [ 67 ] .

*

indicates methods relying on pre-trained data-driven priors.

Neural Implicit Fields
Method	Metric	0000	0059	0106	0169	0181	0207	Avg.
NICE-SLAM [ 103 ]	PSNR $\uparrow$	18.71	16.55	17.29	18.75	15.56	18.38	17.54
	SSIM $\uparrow$	0.641	0.605	0.646	0.629	0.562	0.646	0.621
	LPIPS $\downarrow$	0.561	0.534	0.510	0.534	0.602	0.552	0.548
Vox-Fusion [ 82 ]	PSNR $\uparrow$	19.06	16.38	18.46	18.69	16.75	19.66	18.17
	SSIM $\uparrow$	0.662	0.615	0.753	0.650	0.666	0.696	0.673

Table 15: Rendering performance comparison in PSNR

\uparrow

, SSIM

\uparrow

, and LPIPS

\downarrow

on ScanNet [ 14 ] .

*

indicates methods relying on pre-trained data-driven priors.

Neural Implicit Fields
Method	Rm0	Rm1	Rm2	Off0	Off1	Off2	Off3	Off4	Avg.
NICE-SLAM [ 103 ]	1.69	2.04	1.55	0.99	0.90	1.39	3.97	3.08	1.95
DF-Prior [ 26 ]	1.39	1.55	2.60	1.09	1.23	1.61	3.61	1.42	1.81
Vox-Fusion [ 82 ]	0.27	1.33	0.47	0.70	1.11	0.46	0.26	0.58	0.65
ESLAM [ 33 ]	0.71	0.70	0.52	0.57	0.55	0.58	0.72	0.63	0.63
Point-SLAM [ 59 ]	0.61	0.41	0.37	0.38	0.48	0.54	0.72	0.63	0.52

Table 16: Tracking performance comparisons in ATE RMSE

\downarrow[\mathrm{cm}]

on Replica [ 66 ] .

*

methods relying on pre-trained data-driven priors.

Neural Implicit Fields
Method	fr1/desk	fr2/xyz	fr3/office	Avg.
NICE-SLAM [ 103 ]	4.3	31.7	3.9	13.3
Vox-Fusion [ 82 ]	3.5	1.5	26.0	10.3
Point-SLAM [ 59 ]	4.3	1.3	3.5	3.0
Loopy-SLAM $*$ [ 41 ]	3.8	1.6	3.4	2.9
3D Gaussian Splatting

Table 17: Tracking performance in ATE RMSE

\downarrow[\mathrm{cm}]

on TUM-RGBD [ 67 ] .

*

methods using pre-trained data-driven priors.

Method	fr1/desk	fr2/xyz	fr3/office	Avg.
VTGS-SLAM [ 27 ]	382.4	3462.5	400.5	1415.1
Ours	2.2	1.7	2.0	2.0

Table 18: Tracking performance in ATE RMSE

\downarrow[\mathrm{cm}]

on TUM-RGBD [ 67 ] .

*

indicates VTGS-SLAM [ 27 ] treats each frame as a section, initializing dense Gaussians on each frame similar to our approach.

Method	SplaTAM [ 34 ]	LoopSplat $*$ [ 102 ]	VTGS-SLAM [ 27 ]	GS-ICP SLAM [ 20 ]	Ours(w/o init.)	Ours(w/ init.)
ATE RMSE $\downarrow$ $[\mathrm{cm}]$	11.3	7.3	9.7	115.8	122.4	4.7

Table 19: Tracking performance in ATE RMSE

\downarrow[\mathrm{cm}]

on fr3/nostructure_texture_far in TUM-RGBD [ 67 ] .

*

methods using pre-trained data-driven priors.

Neural Implicit Fields
Method	0000	0059	0106	0169	0181	0207	Avg.
NICE-SLAM [ 103 ]	12.0	14.0	7.9	10.9	13.4	6.2	10.7
Vox-Fusion [ 82 ]	68.8	24.2	8.4	27.3	23.3	9.4	26.9
Point-SLAM [ 59 ]	10.2	7.8	8.7	22.2	14.8	9.5	12.2
Loopy-SLAM $*$ [ 41 ]	4.2	7.5	8.3	7.5	10.6	7.9	7.7
3D Gaussian Splatting

Table 20: Tracking performance in ATE RMSE

\downarrow[\mathrm{cm}]

on ScanNet [ 14 ] .

*

methods using pre-trained data-driven priors.

Method	a	b	c	d	e	Avg.
Neural Implicit Fields
Point-SLAM [ 59 ]	246.16	632.99	830.79	271.42	574.86	511.24
ESLAM [ 33 ]	25.15	2.15	27.02	20.89	35.47	22.14
Loopy-SLAM $*$ [ 41 ]	-	-	25.16	234.25	81.48	113.63
3D Gaussian Splatting
SplaTAM [ 34 ]	1.50	0.57	0.31	443.10	1.58	89.41

Table 21: Tracking performance in ATE RMSE

\downarrow[\mathrm{cm}]

on ScanNet++ [ 83 ] .

*

methods relying on pre-trained data-driven priors.

Method	a	b	c	d	e	Avg.
3D Gaussian Splatting
SplaTAM [ 34 ]	28.02	27.93	29.48	19.65	28.48	26.71
Gaussian-SLAM [ 87 ]	30.06	30.02	31.15	28.75	31.94	30.38
VTGS-SLAM [ 27 ]	32.84	31.02	32.44	31.43	33.38	32.22
LoopSplat $*$ [ 102 ]	30.15	30.08	30.04	28.94	31.78	30.20
Ours	35.95	34.84	35.81	35.71	41.32	36.73

Table 22: Rendering performance comparison in PSNR

\uparrow

on ScanNet++ [ 83 ] .

*

indicates methods relying on pre-trained data-driven priors.

Method	a	b	c	d	e	Avg.
3D Gaussian Splatting
SplaTAM [ 34 ]	23.95	22.66	13.95	8.47	20.06	17.82
Gaussian-SLAM [ 87 ]	26.66	24.42	15.01	18.35	21.91	21.27
VTGS-SLAM [ 27 ]	25.55	24.25	16.94	18.59	21.95	21.46
LoopSplat $*$ [ 102 ]	25.60	23.65	15.87	18.86	22.51	21.30
Ours	26.81	26.79	15.38	20.89	21.71	22.32

Table 23: Novel View Synthesis performance comparison in PSNR

\uparrow

on ScanNet++ [ 83 ] .

*

indicates methods relying on pre-trained data-driven priors. We calculate PSNR including all pixels, regardless of whether they have a valid depth input.

3DGS SLAM

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

See all 63 papers

Geometry-Aware Online Mapping for 3D Gaussian Splatting SLAM
Aug 14, 2026Thai Luu, Quan Tran, Hieu Phan +13DGS SLAM 3D Gaussian Splatting
Cognitive Robotics Lab, Department of Electrical Engineering and Computer Science, University of Arkansas, USA. · Center for AI Research, VinUniversity, Ha Noi, Vietnam.
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
PanoGS-SLAM: Panoramic 3D Gaussian Splatting SLAM
Sep 15, 2026Yongqi Mao, Hao Shi, Yufan Zhang +3Simultaneous Localization and Mapping 3D Gaussian Splatting
Zhejiang University · Ant Group · National University of Defense Technology

Aug 14, 2026J/K move · Enter open · S save

Geometry-Aware Online Mapping for 3D Gaussian Splatting SLAM

Thai Luu, Quan Tran, Hieu Phan, Tuan Dang

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

3DGS SLAM 3D Gaussian Splatting Adaptive Gaussian Densification Streaming 3D Reconstruction

Open paper AskPDF arXiv

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.

Abstract

3DGS SLAM

Explore similar work

Geometry-Aware Online Mapping for 3D Gaussian Splatting SLAM

GS3LAM: Gaussian Semantic Splatting SLAM

PanoGS-SLAM: Panoramic 3D Gaussian Splatting SLAM