Organizations: Graduate School of Information Sciences, Tohoku University, Sendai, Japan. · Department of Electrical Engineering and Information Systems, Graduate School of Engineering, The University of Tokyo, Tokyo, Japan. · Hefei Institutes of Physical Science, Chinese Academy of Sciences, and Department of Automation, University of Science and Technology of China, Hefei, China. · Department of Computer and Systems Engineering, University of Zaragoza, Zaragoza, Spain. · RIKEN Center for AIP, Tokyo, Japan.
RGB-D SLAM requires accurate online tracking, high-quality rendering, and geometrically consistent reconstruction. Existing Gaussian SLAM methods mainly optimize Gaussian primitives with photometric, depth, or local frame-to-model constraints, which are often insufficient in indoor scenes with weak textures, large planar structures, noisy depth, and long trajectories. The resulting maps may suffer from pose drift, surface thickness artifacts, noisy normals, and blurred structural boundaries. We propose DrawSplat, an Atlanta World-guided implicit Gaussian SLAM framework for indoor reconstruction. DrawSplat follows a drawing-like process: explicit point-line geometry first sketches the structural layout, and implicit neural Gaussians then fill in high-fidelity appearance and continuous geometric details. It estimates camera poses, sparse geometry, and gravity direction through point-line joint tracking, infers structural planes from tracked lines using the shared-gravity component of the Atlanta World prior, and converts them into differentiable regularization terms on Gaussian centers and normals. To improve compactness and efficiency, DrawSplat further combines an anchor-based implicit neural Gaussian representation with adaptive structural sampling, which allocates Gaussian capacity to structurally informative regions. Experiments on multiple datasets show that DrawSplat achieves robust tracking, high-fidelity rendering, and strong geometric reconstruction among Gaussian-based SLAM systems, while maintaining a practical balance among quality, efficiency, and model size.
Figures & tables
Fig. 1 : Compared with the cluttered neural Gaussian map of SEGS-SLAM [ 1 ] and the heavy explicit Gaussian map of LoopSplat [ 2 ] , DrawSplat achieves a compact map and a clean mesh by using the Atlanta World assumption to guide implicit neural Gaussians. The right-side radar chart is normalized by using the best and worst values of each metric as boundaries, and shows the relative performance of our method in terms of tracking, rendering, real-time capability, and cost metrics on the Replica [ 3 ] dataset.
Fig. 2 : Overview of DrawSplat. DrawSplat follows a hybrid Gaussian SLAM paradigm [ 36 ] : explicit point-line geometry provides stable tracking and structural cues, while the implicit Gaussian map is incrementally updated online for compact, photorealistic, and geometry-consistent reconstruction. Given an online RGB-D stream, the tracking front-end jointly estimates camera poses, selects keyframes, and builds a sparse 3D geometry map from point and line features. The mapping back-end converts tracked 3D points into a sparse feature grid and decodes implicit Gaussian primitives through lightweight MLPs. Structural 3D lines further induce structural planes constrained by the shared-gravity Atlanta World prior, which regularize Gaussian centers and normals during joint map optimization. The final system produces compact, photorealistic, and structurally consistent indoor reconstructions.
Fig. 3 : Rendering comparison on the Replica Apartment [ 3 ] and ScanNet [ 74 ] datasets. Our method preserves finer visual details.
Fig. 4 : Qualitative geometry reconstruction comparison between our method and representative baselines on the ScanNet dataset [ 74 ] . Our method better preserves scene completeness, structural accuracy, and fine details. The highlighted regions show reconstruction of non-structural objects or locally irregular geometry within otherwise structured indoor scenes, including bicycles, chairs, and table clutter.
Method
Ro0
Ro1
Ro2
Off0
Off1
Off2
Off3
Off4
Avg.
SplaTAM [ 27 ]
0.31
0.47
0.26
0.43
0.24
0.28
0.37
0.58
0.37
Photo-SLAM [ 49 ]
0.49
0.42
0.29
0.56
0.43
1.16
0.80
0.63
0.60
MonoGS [ 28 ]
0.45
0.31
0.29
0.45
0.46
0.25
0.16
1.58
0.49
RTG-SLAM [ 30 ]
0.22
0.27
0.21
0.33
0.20
0.22
0.24
0.44
0.27
DROID-Splat [ 75 ]
0.27
0.28
0.28
0.19
0.25
0.43
0.38
0.34
0.30
Splat-SLAM [ 50 ]
0.33
0.66
0.26
0.30
0.37
0.45
0.41
0.54
0.41
TABLE I : Quantitative comparison of camera tracking accuracy (RMSE [cm] ↓ ) on the simulated Replica dataset [ 3 ] .
Method
0000
0059
0106
0169
0207
Avg.
SplaTAM [ 27 ]
12.57
10.15
17.78
12.65
7.98
12.23
MonoGS [ 28 ]
16.02
12.89
20.18
10.55
11.35
14.20
Photo-SLAM [ 49 ]
7.88
7.85
9.84
9.96
7.53
8.61
RTG-SLAM [ 30 ]
8.11
6.75
9.03
10.24
9.33
8.69
MG-SLAM* [ 31 ]
5.95
6.41
8.07
7.29
6.14
6.77
LoopSplat [ 2 ]
5.85
6.77
7.16
9.58
6.82
7.24
TABLE II : Quantitative comparison of camera tracking accuracy (RMSE [cm] ↓ ) on the real-world ScanNet dataset [ 74 ] .
Method
apart0
apart1
apart2
frl0
frl4
Avg.
SplaTAM [ 27 ]
25.14
15.67
6.17
2.74
3.31
10.61
Photo-SLAM [ 49 ]
5.39
7.98
6.31
1.09
1.27
4.41
MonoGS [ 28 ]
20.69
9.51
7.57
2.98
6.14
9.38
RTG-SLAM [ 30 ]
6.99
9.09
6.90
1.33
0.98
5.06
MG-SLAM* [ 31 ]
5.03
3.87
3.15
0.92
0.88
2.77
LoopSplat [ 2 ]
5.61
21.12
0.64
0.60
0.87
5.77
TABLE III : Quantitative comparison of camera tracking accuracy (RMSE [cm] ↓ ) on the simulated Replica Apartment dataset [ 3 , 78 ] .
Fig. 5 : Qualitative geometry reconstruction comparison between our method and representative baselines on the ScanNet++ dataset [ 77 ] . Our method achieves smoother surfaces.
Method
Metric
Ro0
Ro1
Ro2
Off0
Off1
Off2
Off3
Off4
Avg.
Spla TAM [ 27 ]
PSNR ↑
32.84
33.81
35.44
38.52
38.72
31.78
30.15
31.55
34.10
SSIM ↑
0.968
0.965
0.970
0.981
0.975
0.955
0.943
0.941
0.962
LPIPS ↓
0.072
0.095
0.071
0.083
0.098
0.099
0.118
0.153
0.099
Mono GS [ 28 ]
PSNR ↑
34.57
36.05
37.57
42.11
42.03
36.33
36.38
36.16
37.65
SSIM ↑
0.961
0.952
0.973
0.980
0.976
0.972
0.969
0.965
0.969
LPIPS ↓
0.085
0.088
0.079
0.053
0.064
0.068
0.079
0.088
0.076
TABLE IV : Quantitative comparison of appearance rendering performance on the Replica dataset [ 3 ] .
Method
Metric
0000
0059
0106
0169
0207
Avg.
SplaTAM [ 27 ]
PSNR ↑
17.95
20.11
19.85
20.09
17.98
19.20
SSIM ↑
0.611
0.796
0.739
0.795
0.710
0.730
LPIPS ↓
0.472
0.291
0.333
0.261
0.346
0.341
MonoGS [ 28 ]
PSNR ↑
20.28
20.05
20.11
20.56
20.95
20.39
SSIM ↑
0.773
0.768
0.785
0.792
0.802
0.784
LPIPS ↓
0.385
0.295
0.273
0.251
0.293
0.299
TABLE V : Quantitative comparison of appearance rendering performance on the ScanNet dataset [ 74 ] .
Method
Metric
apart0
apart1
apart2
frl0
frl4
Avg.
SplaTAM* [ 27 ]
PSNR ↑
13.12
24.57
28.52
31.82
32.71
26.15
SSIM ↑
0.415
0.821
0.883
0.930
0.945
0.799
LPIPS ↓
0.656
0.302
0.241
0.164
0.184
0.309
Photo-SLAM* [ 49 ]
PSNR ↑
29.07
22.73
24.59
34.16
33.36
28.78
SSIM ↑
0.922
0.796
0.848
0.940
0.932
0.888
LPIPS ↓
0.227
0.293
0.354
0.115
0.129
0.224
TABLE VI : Quantitative comparison of appearance rendering performance on the Replica Apartment dataset [ 3 , 78 ] . * represents the baseline results of [ 27 , 49 , 28 , 30 , 31 ] come from MG-SLAM [ 31 ] . The dash indicates system failures.
Method
Metric
Ro0
Ro1
Ro2
Off0
Off1
Off2
Off3
Off4
Avg.
Neural Implicit Fields
NICE-SLAM [ 16 ]
Depth L1 ↓
1.75
1.34
1.99
1.45
1.65
6.95
4.82
2.31
2.78
F1-score ↑
90.2
91.9
92.2
93.1
82.2
88.0
87.1
87.7
89.1
Co-SLAM [ 17 ]
Depth L1 ↓
1.05
0.85
2.37
1.24
1.48
1.86
1.66
1.54
1.51
ESLAM [ 18 ]
Depth L1 ↓
0.99
1.21
1.17
0.97
1.19
1.62
1.55
1.01
1.21
F1-score ↑
94.8
95.0
95.7
97.5
95.8
93.1
93.9
92.2
94.7
TABLE VII : Quantitative comparison of geometry reconstruction accuracy on the Replica dataset [ 3 ] .
Method
0000
0059
0106
0169
0207
Avg.
SplaTAM [ 27 ]
7.87
5.63
7.13
5.28
6.40
6.46
MonoGS [ 28 ]
56.67
50.95
47.47
28.10
54.41
47.52
Splat-SLAM [ 50 ]
8.21
15.54
18.43
11.31
10.54
12.81
LoopSplat [ 2 ]
10.93
25.16
45.42
45.72
17.18
28.88
Ours
3.05
5.06
4.76
5.26
3.91
4.41
TABLE VIII : Quantitative comparison of 2D geometry reconstruction accuracy (Depth L1 ↓ ) on the ScanNet dataset [ 74 ] .
Method
PSNR [dB] ↑
Track. FPS ↑
Render. FPS ↑
SLAM FPS ↑
Memory [GB] ↓
Size [MB] ↓
SplaTAM [ 27 ]
34.10
0.26
121
0.13
18.8
295.5
Photo-SLAM [ 49 ]
34.86
28.87
494
13.02
7.3
61.75
MonoGS [ 28 ]
37.65
2.11
465
1.28
16.5
44.23
LoopSplat [ 2 ]
36.58
0.68
138
0.35
9.8
73.28
SEGS-SLAM [ 1 ]
39.32
15.29
143
1.73
4.7
78.98
Ours
39.43
24.55
209
2.02
4.5
18.29
TABLE IX : System performance comparison. The reported values are averaged over all sequences in the Replica dataset [ 3 ] . Experiments are conducted on an NVIDIA A100 GPU and an AMD EPYC 7542 CPU. All images are processed at their original resolution. DrawSplat achieves strong reconstruction quality with a compact model, real-time tracking, and fast rendering.
Method
apart0
scene0000
ATE ↓
PSNR ↑
Track. FPS ↑
ATE ↓
PSNR ↑
Track. FPS ↑
(a)
Only point
8.12
34.86
34.72
6.25
24.18
41.36
(b)
(a) + Line
3.81
37.42
25.64
5.48
26.31
31.75
(c)
(a) + Global BA
5.35
36.05
30.18
5.72
25.09
36.42
(d)
(b) + Global BA
3.21
38.21
23.16
5.03
26.86
29.37
(e)
(d) + Point planes + point gravity
3.21
38.03
21.74
5.03
26.54
26.83
TABLE X : Ablation study of the tracking front-end on Replica Apartment [ 3 , 78 ] apart0 and ScanNet [ 74 ] scene0000. ATE RMSE ↓ is reported in [cm].
Method
PSNR ↑
F1-score ↑
Render. FPS ↑
Gaussians ↓
Size [MB] ↓
Our Track. + 3DGS [ 22 ]
40.57
-
346
437231
108.43
Our Track. + GOF* [ 58 ]
39.05
89.57
279
1036097
261.10
voxel_size = 0.001
40.48
92.66
152
127712
54.15
voxel_size = 0.01 ( ours )
39.43
93.71
209
53380
18.29
voxel_size = 0.1
35.92
92.62
238
6529
2.77
TABLE XI : Ablation study of the representation design. The reported values are averaged over all sequences in the Replica dataset [ 3 ] . * represents that we further add depth optimization to GOF [ 58 ] .
Fig. 6 : Ablation study of Atlanta World-guided structural regularization on ScanNet dataset [ 74 ] . The proposed regularization reduces sensor noise and produces smoother geometric structures.
Method
PSNR ↑
F1-score ↑
Size [MB] ↓
Render. FPS ↑
Gradient-based densification
38.61
91.92
16.95
217
w/o structural weight
39.02
92.76
16.82
218
w/o line response
39.12
92.87
17.60
212
w/o depth-gradient response
39.21
93.10
17.74
211
w/o photometric residual
39.13
93.21
18.11
213
w/o pruning
39.55
93.84
23.17
166
TABLE XII : Ablation study of the proposed adaptive structural sampling strategy on the Replica dataset [ 3 ] .
Source
PSNR ↑
F1-score ↑
mIoU w ↑
IoU f ↑
IoU c ↑
FPS ↑
GT sem.
39.48
93.81
0.989
0.954
0.836
-
Predicted sem.
39.43
93.71
0.986
0.964
0.809
9.2
TABLE XIII : Ablation study of semantic source robustness on the Replica dataset [ 3 ] .
Method
PSNR ↑
Depth L1 ↓
F1-score ↑
Basic
39.85
3.86
73.39
w/o Ld
38.02
2.97
77.22
w/o Ln
39.23
0.68
93.44
w/o Lgp
39.08
0.64
93.31
w/o Lgn
39.25
0.58
93.50
w/o Law
39.33
0.54
93.55
TABLE XIV : Ablation study of the map loss and semantic weighting design on the Replica dataset [ 3 ] .
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.
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.
3D Gaussian Splatting (3DGS) has significantly boosted novel view synthesis and high-fidelity scene reconstruction, expanding the potential of 3DGS-based Visual Simultaneous Localization and Mapping (SLAM) methods. However, most existing systems fail to fully exploit the underlying structural information, which limits rendering quality and often leads to inconsistent maps. To address these limitations, we propose MMD-SLAM, a structure-enhanced Visual SLAM framework that leverages the Atlanta World (AW) assumption to guide a Multi-Meta Gaussian representation for photorealistic mapping. First, we introduce a point-line fusion strategy for pose optimization, where 3D line segments are incorporated to improve tracking robustness and provide additional constraints for mapping. Second, we design a Multi-Meta Gaussian representation with dominant directions, explicitly encoding structural priors from the AW hypothesis. Finally, we propose a Gaussian evolution strategy that adapts to scene geometry and incorporates structural cues into global optimization. Extensive experiments demonstrate that these innovations enable MMD-SLAM to achieve state-of-the-art performance in both tracking accuracy and mapping quality. e.g., our method achieves a 48.56% reduction in ATE RMSE on ScanNet and a 5.71% improvement in PSNR on Replica, compared with MonoGS.
Fan Zhu, Ziyu Chen, Peichen Liu +6
HFIPS, Chinese Academy of Sciences, Hefei, China. · University of Science and Technology of China, Hefei, China. · Aarhus University, Aarhus, Denmark. +3
SLAM systems based on 3D Gaussian Splatting (3DGS) have recently demonstrated promising reconstruction accuracy for dense 3D scene representations. However, current 3DGS systems struggle to meet the strict demands of real-world deployments due to severe limitations in operational performance and map adaptability. To this end, we propose LightSplat, a hybrid-representation RGB-D SLAM framework. It synergizes local sparse features for robust and fast tracking with a dual-thread backend that progressively constructs dense Gaussian submaps. Crucially, we enable online loop closure through feature-accelerated 3DGS registration, refining overall map consistency through pose graph optimization. Ultimately, LightSplat achieves the online reconstruction of high-fidelity Gaussian map. Extensive experiments on multiple datasets and real-world robotic platform demonstrate that our method achieves near state-of-the-art reconstruction quality and the capability to accommodate practical camera motions, maintaining an average framerate of 8 FPS. Overall, LightSplat provides an efficient and robust foundation for deploying high-fidelity 3DGS in real-world environments.
Junze Bao, Ye Gao, Yiming Huang +5
School of Automation Science and Electrical Engineering, Beihang University, Beijing 100191, P. R. China, and also with Zhejiang Key Laboratory of Industrial Big Data and Robot Intelligent Systems, Hangzhou Innovation Institute, Beihang University, Hangzhou 310051, P. R. China. · School of Mechanical Engineering, Zhejiang Sci-Tech University, Hangzhou 310018, P. R. China.