Organizations: University of Science and Technology of China, Hefei, China · National Key Laboratory of Deep Space Exploration, Deep Space Exploration Laboratory, Hefei, China
3D Gaussian Splatting has achieved remarkable success in novel view synthesis. However, extracting high-fidelity surfaces directly from 3DGS remains challenging due to its discrete and unstructured nature. Existing 3DGS-based reconstruction methods typically rely on multi-view geometric consistency or local constraints. Without an explicit structured geometric prior during optimization, these methods often struggle to resolve structural ambiguities, leading to artifacts and floaters, particularly in textureless or occluded regions. To address this limitation, we propose TopoSurfel, a novel framework that closes the loop between Gaussian surfels and continuous meshes. Unlike recent methods that incorporate mesh extraction into the differentiable pipeline by introducing auxiliary neural networks or extra per-Gaussian parameters, we dynamically extract a continuous proxy mesh via a non-trainable differentiable iso-surfacing process. Leveraging this differentiable connection, we introduce a mesh-guided surfel evolution strategy, including normal alignment and geometry-aware density control, to effectively suppress floaters and fill surface holes. Furthermore, to address the initialization challenges in large-scale environments, we propose a spatially aware hybrid re-initialization strategy that ensures robust reconstruction across complex scenes. Extensive experiments demonstrate that TopoSurfel achieves competitive geometric reconstruction accuracy while maintaining high-quality mesh-based novel view synthesis. The code for our method is available at https://github.com/Fan-Treasure/TopoSurfel.
Figures & tables
Figure 1. Overview of our surface reconstruction framework. (a) We propose a synergistic co-evolution scheme that bridges Gaussian surfels and differentiable meshes. By enforcing mutual geometry guidance, our method effectively captures both high-frequency details and coherent surface structures. (b) Reconstruction results on various scenes demonstrate that our method outperforms existing methods in both quality and efficiency. Overview of our surface reconstruction framework.
Figure 2. Overview of the TopoSurfel Pipeline. (Left) Warm-up: We initialize Gaussian surfels from SfM points and obtain an initial mesh via TSDF fusion. (Right) Closed-Loop Optimization: In addition to the surfel training pipeline, we extract a weighted oriented point cloud from surfels, and then reconstruct the Differentiable Proxy Mesh from it using DPSR and DiffMC. Both the proxy mesh and the surfels are rendered to compute geometry consistency losses. Overview of the TopoSurfel Pipeline.
Figure 3. Illustration of Our Mesh-Guided Normal Alignment and Geometry-Aware Density Control Illustration of mesh-guided normal alignment and geometry-aware density control.
CD (mm) ↓
24
37
40
55
63
65
69
83
97
105
106
110
114
118
122
Avg.
Time
NeRF- Based
VolSDF ( Yariv et al., 2021 )
1.14
1.26
0.81
0.49
1.25
0.70
0.72
1.29
1.18
0.70
0.66
1.08
0.42
0.61
0.55
0.86
> 12h
NeuS ( Wang et al., 2021 )
0.83
0.98
0.56
0.37
1.13
0.59
0.60
1.45
0.95
0.78
0.52
1.43
0.36
0.45
0.45
0.77
> 12h
Neuralangelo ( Li et al., 2023 )
0.45
0.74
0.33
0.34
1.05
0.54
0.53
1.33
1.05
0.72
0.43
0.69
0.34
0.38
0.42
0.62
> 12h
Gaussian- Based
SuGaR ( Su et al., 2023 )
1.47
1.33
1.13
0.61
2.25
1.71
1.15
1.63
1.62
1.07
0.79
2.45
0.98
0.88
0.79
1.33
1h
2DGS ( Huang et al., 2024 )
0.46
0.84
0.31
0.45
0.92
1.01
0.83
1.23
1.30
0.66
0.61
1.07
0.45
0.71
0.54
0.76
11m
GOF ( Yu et al., 2024b )
0.50
0.82
0.37
0.37
1.12
0.74
0.73
1.18
1.29
0.68
0.77
0.90
0.42
0.66
0.49
0.74
1h
Table 1. Quantitative geometry comparison on the DTU scenes. Our method achieves the lowest Chamfer Distance without monocular depth priors, while maintaining competitive efficiency. denotes the best, second best, third best, respectively. † indicates the use of an additional monocular depth prior.
F1-Score ↑
Geo-NeuS
N-angelo
RaDe-GS
PGSR
QGS
MILo (dense)
Ours
Barn
0.33
0.70
0.43
0.66
0.55
0.64
0.65
Caterpillar
0.26
0.36
0.32
0.41
0.40
0.38
0.43
Courthouse
0.12
0.28
0.21
0.21
0.28
0.31
0.22
Ignatius
0.72
0.89
0.69
0.80
0.81
0.76
0.77
Meetingroom
0.20
0.32
0.25
0.29
0.31
0.28
0.34
Truck
0.45
0.48
0.51
0.60
0.64
0.59
0.64
Table 2. Quantitative geometry comparison on the TNT dataset.
Indoor scenes
Outdoor scenes
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
NeRF
26.84
0.790
0.370
21.46
0.458
0.515
Deep Blending
26.40
0.844
0.261
21.54
0.524
0.364
I-NGP
29.15
0.880
0.216
22.90
0.566
0.371
Mip-NeRF 360
31.72
0.917
0.180
24.47
0.691
0.283
3DGS
30.52
0.921
0.199
24.45
0.728
0.240
Table 3. Quantitative comparison of NVS on the Mip-NeRF 360 dataset.
Figure 4. Qualitative comparison of surface reconstruction on the DTU dataset. In heavily occluded regions and areas with few viewpoints, our method recovers more complete geometry and smoother surfaces with fewer artifacts, whereas other methods often produce over-smoothed or noisy surfaces. Qualitative geometry comparison on DTU scenes.
Figure 5. Qualitative results on the TNT dataset. Our method excels in large-scale real-world scenes, producing cleaner surfaces and more accurate structures. While methods like MILo and QGS struggle with holes or floating noise in complex areas, TopoSurfel maintains high geometric fidelity. Qualitative geometry comparison on TNT scenes.
Figure 6. Novel view synthesis results on the Mip-NeRF 360 dataset. TopoSurfel achieves comparable visual quality to existing methods. NVS comparison on mipnerf360 scenes.
2DGS
GOF
RaDe-GS
PGSR
IMLS-Splat
MILo
Ours
PSNR ↑
22.10
23.97
23.58
23.14
23.72
23.77
25.23
SSIM ↑
0.886
0.905
0.903
0.900
0.906
0.910
0.921
LPIPS ↓
0.106
0.092
0.094
0.095
0.095
0.094
0.087
∣V∣
0.4M
4.1M
3.0M
3.2M
0.3M
1.0M
1.7M
Table 4. Quantitative comparison of mesh-based Novel View Synthesis on the NeRF-Synthetic dataset.
Figure 7. Qualitative comparison of Mesh-based NVS on the NeRF-Synthetic dataset. As shown in the zoomed-in regions, our approach preserves intricate details that are often blurred or lost in other methods. Mesh-based novel view synthesis results on NeRF-Synthetic.
Training Resources
Rendering
Output Mesh
#GS (M)
GPU Mem
Time
FPS
#Vertices
Size (MB)
2DGS
0.10
4.1 GB
11m
490
0.4 M
15.32
GOF
0.18
10.1 GB
1h
41
3.8 M
277.1
RaDe-GS
0.13
11.3 GB
14m
436
2.8 M
143.7
PGSR
0.16
6.4 GB
30m
227
3.0 M
152.5
MILo
0.20
9.4 GB
43m
114
0.7M
41
Table 5. Comparison of training resources, rendering performance, and output mesh size across methods, where #GS denotes the number of Gaussians (in millions), GPU Mem denotes peak GPU memory usage (GB), Time denotes training time, FPS denotes frames per second, #Vertices denotes the number of mesh vertices, and Size denotes mesh file size (MB).
Dataset
Grid Resolution
Peak Memory
Time
CD/F1
DTU
2563
4.2 GB
27m
0.54
5123
8.5 GB
37m
0.51
7203
17.5 GB
60m
0.50
TNT
2563
9.6 GB
76m
0.45
5123
14.7 GB
98m
0.52
7203
OOM
–
–
Table 6. Memory stress test under different DPSR grid resolutions on an RTX 3090 GPU with 24 GB memory. We report peak memory, training time, Chamfer distance on DTU ( ↓ ), and F1-score on TNT ( ↑ ).
Setting
F1-Score ↑
PSNR ↑
w/o Mesh Loop
0.42
26.75
w/o Warm-up
0.39
26.46
View-Statistics Normal Flips
0.46
26.91
Random Normal Flips
0.37
26.20
w/o Hybrid Initialization
0.21
20.59
w/o Geo-Aware Density Control
0.48
27.03
Table 7. Ablation study on the TNT dataset.
Figure 8. Ablation study on the warm-up stage. (a) Gaussian point clouds colored by depth and (b) extracted proxy meshes. In each subfigure, the left side shows the result without warm-up, while the right side shows the result with our warm-up initialization. Ablation of the warm-up stage and mesh-in-the-loop optimization.
Figure 9. Ablation study on mesh-guided normal alignment. (a) Gaussian point clouds colored by surfel normals and (b) extracted proxy meshes. For both (a) and (b), we compare the baseline using only view statistics (left) against our full strategy with mesh-guided normal alignment (right). Ablation of mesh-guided normal alignment.
Figure 10. Ablation study on hybrid re-initialization for large-scale scenes. (a) Without hybrid re-initialization. (b) With hybrid re-initialization. From left to right, we show the rendered RGB images, normal maps, and depth maps. Ablation of hybrid re-initialization for large-scale scenes.
Figure 11. Ablation study on geometry-aware density control. (a) Gaussian point clouds colored by depth and (b) final reconstructed meshes. Within each subfigure, the left side uses only gradient-guided density control, whereas the right side incorporates our geometry-aware density control. Ablation of geometry-aware density control.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Hyperparameter
Value
Optimization hyperparameters
position_lr_init
1.6×10−4
position_lr_final
1.6×10−6
feature_lr
2.5×10−3
opacity_lr
5.0×10−2
scaling_lr
5.0×10−3
Appendix
Table 8. Hyperparameters used in our implementation.
Setting
F1-Score ↑
PSNR ↑
Time
Mesh extraction
DPSR final Extraction
0.43
26.83
98m
FlexiCubes Extraction
0.39
26.67
137m
Sampling and weighting
1-point + opacity
0.49
26.85
92m
1-point + opacity-area
0.51
26.81
96m
Appendix
Table 9. Additional ablation studies on the TNT dataset.
Figure 12. Ablation of final mesh extraction with DPSR and FlexiCubes. (a) Mesh extracted with DPSR. (b) Mesh extracted with FlexiCubes. (c) Mesh extracted with TSDF fusion. Final Mesh Extraction with DPSR and FlexiCubes.
Figure 13. Ablation of normal estimation without Gaussian flattening (a) Mesh extracted with GOF’s normal estimation. (b) Mesh extracted with RaDe-GS’s normal estimation. (c) Mesh extracted with our normal estimation. Normal Estimation without Gaussian Flattening.
Figure 14. Visualization of Our Reconstructed Meshes on the DTU Dataset. Visualization of Our Reconstructed Meshes on the DTU Dataset.
Method
Chair
Drums
Ficus
Hotdog
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
2DGS
25.84
0.915
0.066
20.27
0.881
0.108
17.68
0.876
0.120
26.33
0.933
0.071
GOF
26.50
0.935
0.053
21.54
0.896
0.101
23.63
0.923
0.083
27.32
0.950
0.062
PGSR
26.92
0.935
0.055
20.64
0.888
0.100
22.82
0.905
0.082
27.22
0.953
0.057
IMLS-Splatting
27.10
0.943
0.047
21.53
0.901
0.100
23.64
0.916
0.088
28.18
0.958
0.053
Ours
28.01
0.943
0.056
22.50
0.914
0.085
25.19
0.940
0.068
30.52
0.967
0.051
Appendix
Table 10. Quantitative evaluation of novel view synthesis on each scene of the NeRF-Synthetic dataset.
Figure 15. Visualization of Our Reconstructed Meshes and trained UV map on the NeRF-Synthetic Dataset. Visualization of Our Reconstructed Meshes and trained UV map on the NeRF-Synthetic Dataset.
Figure 16. Evolution of the differentiable proxy mesh during closed-loop optimization. The proxy mesh is used as a training-time geometric prior. Iterations are counted from the beginning of the mesh-guided stage. Evolution of the differentiable proxy mesh during closed-loop optimization.
Figure 17. Failure case on the Materials scene of NeRF-Synthetic. Failure Case.
Figure 18. Comprehensive visualization of results on multiple Mip-NeRF 360 scenes, including ground-truth images, our rendered images, reconstructed meshes, Surfel normals, and mesh normals. Comprehensive visualization of results on multiple Mip-NeRF 360 scenes, showing rendered images, reconstructed meshes, Surfel normals, and mesh normals.
3D Gaussian Splatting (3DGS) has recently enabled real-time novel view synthesis with impressive quality. However, it struggles to recover accurate surfaces under limited viewpoints and due to the inherent irregularity of Gaussian primitives. The resulting geometric errors are notoriously difficult to correct manually. To address these issues, we propose Gaussian Sculpting, a fully differentiable end-to-end framework for high-quality surface reconstruction. Our key insight is to anchor Gaussians onto an evolving differentiable surface, allowing them to guide signed distance field (SDF) optimization instead of extracting the surface only during post-processing. To enable stable gradient isolation during joint optimization, we design a bi-level training strategy in which the outer loop optimizes the geometry represented by the SDF, while the inner loop updates the Gaussians with the geometry fixed. We further impose constraints on Gaussian parameters to ensure consistency with the underlying surface, thereby improving both geometric and appearance fidelity during optimization. In addition, we introduce a multi-resolution subdivision scheme based on octree-like partitioning to preserve fine details while reducing memory consumption. Experiments on object-level scenes demonstrate that our method effectively removes redundant surfaces, recovers missing structures caused by limited viewpoints, and achieves strong reconstruction quality even at relatively low resolutions.
Ke Jiaxin, Juncheng Liu, Yi Wang +4
Dalian University of Technology, China · Massey University, New Zealand · Peking University, China +1
3D Gaussian Splatting (3DGS) has emerged as a powerful technique for generating photorealistic renderings of a scene in real-time. However, the volumetric nature of 3DGS limits its ability to accurately capture surface geometry. To address this, 2D Gaussian Splatting (2DGS) was proposed to enable view-consistent and geometrically accurate surface reconstruction from multi-view images. However, 2DGS can be sensitive to the initialization of the Gaussian primitives. Reliance on Structure-from-Motion (SfM) initializations, which can produce poor estimates on challenging image sets, may lead to subpar results. In this work, we enhance 2DGS by incorporating monocular depth and normal priors to improve both geometric accuracy and robustness. We propose a depth-guided initialization strategy for Gaussians and introduce a clustering-based technique for pruning degenerate Gaussians. We evaluate our method on the DTU dataset, where it achieves state-of-the-art results in mesh reconstruction while preserving high-quality novel view synthesis.
Prajwal Gupta C. R., Divyam Sheth, Jinjoo Ha +2
TU Darmstadt · ELIZA · Max Planck Institute for Intelligent Systems
High-fidelity surface reconstruction from multi-view images is a core problem in 3D computer vision. While neural implicit surfaces like SDFs offer smooth geometry, they are often bottlenecked by the computational intensity of volume rendering. Conversely, 3D Gaussian Splatting (3DGS) provides rapid training but lacks geometry continuity, often leading to fragmented surfaces. This paper presents a novel framework that integrates Signed Distance Fields directly into the splatting pipeline. By leveraging the continuous nature of SDFs to regularize Gaussian primitives, our method effectively fills geometric holes and suppresses noise inherent in sparse point clouds. Unlike hybrid approaches that rely on heavy volumetric sampling, our approach utilizes the efficiency of splatting to achieve faster convergence. Extensive evaluations demonstrate that our method produces high-quality surfaces with significantly fewer primitives, offering a more compact and efficient representation for both indoor and outdoor environments.
Baixin Xu, Jiangbei Hu, Jiaze Li +1
College of Computing and Data Science, Nanyang Technological University · School of Software, Dalian University of Technology