Confidence-Based Mesh Extraction from 3D Gaussians

Lukas Radl; Felix Windisch; Andreas Kurz; Thomas Köhler; Michael Steiner; Markus Steinberger

cs.CVMar 25, 2026

Confidence-Based Mesh Extraction from 3D Gaussians

Authors: Lukas Radl, Felix Windisch, Andreas Kurz, Thomas Köhler, Michael Steiner, Markus Steinberger

Organizations: Graz University of Technology, Austria · Huawei Technologies, Austria

Abstract

Recently, 3D Gaussian Splatting (3DGS) greatly accelerated mesh extraction from posed images due to its explicit representation and fast software rasterization. While the addition of geometric losses and other priors has improved the accuracy of extracted surfaces, mesh extraction remains difficult in scenes with abundant view-dependent effects. To resolve the resulting ambiguities, prior works rely on multi-view techniques, iterative mesh extraction, or large pre-trained models, sacrificing the inherent efficiency of 3DGS. In this work, we present a simple and efficient alternative by introducing a self-supervised confidence framework to 3DGS: within this framework, learnable confidence values dynamically balance photometric and geometric supervision. Extending our confidence-driven formulation, we introduce losses which penalize per-primitive color and normal variance and demonstrate their benefits to surface extraction. Finally, we complement the above with an improved appearance model, by decoupling the individual terms of the D-SSIM loss. Our final approach delivers state-of-the-art results for unbounded meshes while remaining highly efficient.

Figures & tables

Panel of Figure 2 — Figure 2: Analysis of the Photometric Loss Components : The luminance term of D-SSIM is highly dependent on the illumination; in this example, the luminance error is reduced by a factor of $9\times$ by considering the appearance embedding.

Panel of Figure 3 — Figure 3: Confidence-Driven Gaussian Splatting: We show rendered images from our trained method, accompanied by the rendered confidence maps $\hat{C}$ . The confidence maps effectively isolate reflective surfaces, thin foliage, or rarely observed areas (such as the roof for Barn). Importantly, the low-confidence regions are still well-reconstructed.

Panel of Figure 4 — Figure 4: Effect of our proposed Variance Losses: Rendering color and normals only for the first-hit Gaussians shows how our variance losses align individual Gaussians better to the true object surface, in both color and orientation.

	Barn	Caterpillar	Courthouse	Ignatius	Meetingroom	Truck	Average	Runtime
PGSR	0.548	0.437	0.238	0.728	0.367	0.658	0.496	28min
QGS	0.536	0.374	0.183	0.733	0.374	0.645	0.474	41min
GOF	0.484	0.402	0.288	0.674	0.275	0.596	0.453	40min
SOF	0.535	0.408	0.297	0.736	0.309	0.558	0.474	17min
MILo	0.541	0.389	0.322	0.757	0.281	0.617	0.485	60min
Ours	0.534	0.472	0.333	0.782	0.372	0.634	0.521	18min

Table 1: Full Geometry Reconstruction results for the Tanks & Temples dataset [ 31 ] . We evaluate the F1-score (higher is better) and report runtimes, measured on an RTX 4090. Overall, our method achieves the best reconstruction quality of all tested methods (without leveraging multi-view constraints), while remaining efficient.

	5a269ba6fe	08bbbdcc3d	39f36da05b	c263dfbf0	ef18cf0708	fb564c935d	Average
PGSR	0.668	0.666	0.647	0.630	0.542	0.632	0.631
QGS	0.646	0.622	0.571	0.522	0.504	0.572	0.573
GOF	0.620	0.670	0.611	0.672	0.504	0.661	0.623
SOF	0.599	0.681	0.616	0.675	0.497	0.621	0.615
MILo	0.622	0.675	0.621	0.626	0.542	0.660	0.624
Ours	0.670	0.729	0.657	0.715	0.551	0.684	0.668

Table 2: Full Geometry Reconstruction results for our ScanNet++-v2 subset [ 69 ] , where we report the F1-score. Our method attains the highest scores across all scenes, underscoring our real-world applicability.

	Tanks & Temples			ScanNet++
Method	Precision $\uparrow$	Recall $\uparrow$	F1-Score $\uparrow$	Precision $\uparrow$	Recall $\uparrow$	F1-Score $\uparrow$
SOF [ 52 ] (baseline)	0.542	0.438	0.474	0.565	0.677	0.615
+ Improved Appearance	0.557	0.459	0.493	0.577	0.685	0.625
+ ${\mathcal{L}}_{\text{conf}}$	0.568	0.479	0.509	0.605	0.715	0.655
+ ${\mathcal{L}}_{\text{color-var}}$	0.571	0.491	0.519	0.606	0.722	0.658
+ ${\mathcal{L}}_{\text{normal-var}}$	0.580	0.491	0.521	0.613	0.734	0.668

Table 3: Ablation Study for our Method: We analyze the effects of our individual contribution on both the Tanks & Temples [ 31 ] and the ScanNet++ dataset [ 69 ] . As we can see, all our proposed components result in notable improvements.

Panel of Figure 5 — Figure 5: Qualitative Mesh Comparison : Our approach produces more detailed meshes with fewer artifacts than other unbounded methods [ 18 , 52 ] . Additionally, it achieves higher completeness and finer detail than bounded extraction works [ 5 , 77 ] .

Panel of Figure 6 — Figure 6: Unbounded Mesh Rendering: Our method reconstructs both fine details and high-quality backgrounds on challenging Mip-NeRF 360 scenes [ 1 ] .

	Baselines (Standard)				Ours (Ablations)
Method	Precision $\uparrow$	Recall $\uparrow$	F1-Score $\uparrow$	Method	Precision $\uparrow$	Recall $\uparrow$	F1-Score $\uparrow$
VastG. [ 37 ]	0.541	0.440	0.475	VastG. (Imp.)	0.557	0.445	0.484
PGSR [ 5 ]	0.545	0.441	0.478	w/o $l(\cdot)$	0.559	0.441	0.483
H3DGS [ 29 ]	0.553	0.448	0.484	w/o SSIM Dec.	0.559	0.453	0.490
PPISP [ 10 ]	0.556	0.456	0.489	Ours	0.557	0.459	0.493

Table 4: Ablation Study on the Decoupled Appearance Module: We compare appearance embeddings via surface reconstruction on the Tanks & Temples dataset [ 31 ] . Our proposed decoupled configuration yields the strongest results.

Panel of Figure 7 — Figure 7: $\beta$ -Ablation Study: We evaluate surface reconstruction quality for varying choices of $\beta$ . Our chosen value of $7.5\times 10^{-2}$ achieves the best performance.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

	Barn	Caterpillar	Courthouse	Ignatius	Meetingroom	Truck	Average	# Gaussians
Ours	0.534	0.472	0.333	0.782	0.372	0.634	0.521	1.27M
Ours w/o adapted densification	0.537	0.463	0.338	0.769	0.369	0.623	0.516	1.53M
Ours	0.534	0.472	0.333	0.782	0.372	0.634	0.521	1.27M
Ours w/ ${\mathcal{L}}_{\text{conf}}$ from iter. 2.5K	0.532	0.465	0.330	0.783	0.363	0.637	0.518	1.34M
Ours w/ ${\mathcal{L}}_{\text{conf}}$ from iter. 5K	0.537	0.465	0.323	0.780	0.366	0.632	0.517	1.42M
Ours w/ ${\mathcal{L}}_{\text{conf}}$ from iter. 10K	0.537	0.447	0.328	0.764	0.359	0.616	0.508	1.43M

Appendix

Table 5: Extended Ablation Studies for the Tanks & Temples dataset [ 31 ] . We investigate the impact of our confidence-steered densification, our choice of using

{\mathcal{L}}_{\text{conf}}

from iteration

500

, and verify our chosen hyperparameters.

Panel of Figure 8 — Figure 8: ScanNet++ dataset [ 69 ] : We show example images from our selected scenes, with the number of images per-scene inset. All selected scenes contain challenging reflections ( e.g . 5a269ba6fe, fb564c935d ), textureless areas, and motion blur.

	Barn	Caterpillar	Courthouse	Ignatius	Meetingroom	Truck	Average	Runtime
Ours	0.534	0.472	0.333	0.782	0.372	0.634	0.521	18min
GeoSVR	0.557	0.495	0.347	0.804	0.382	0.688	0.546	42min
GeoSVR (w/ QGS Eval)	0.544	0.432	0.328	0.787	0.408	0.651	0.525	42min

Appendix

Table 6: Comparison with prior-dependent GeoSVR [ 33 ] using the Tanks & Temples dataset [ 31 ] . Our method is able to extract unbounded meshes while keeping processing time low, whereas GeoSVR attains a higher F1-score.

Panel of Figure 9 — Figure 9: Completeness Comparison: Due to gaps in the ground-truth point clouds, our more complete meshes are penalized in precision, compared to the bounded meshes from PGSR [ 5 ] (black points denote low precision). This highlights the fundamental disadvantage unbounded methods face for this benchmark.

Panel of Figure 10 — Figure 10: Novel View Mesh Rendering Comparison with Bounded Methods. Our meshes are significantly more complete and detailed compared to bounded methods, such as PGSR [ 5 ] .

Panel of Figure 11 — Figure 11: Novel View Mesh Rendering Comparison with Unbounded Methods. Our method consistently preserves geometry and intricate topological details that MILo [ 18 ] smooths over or misrepresents.

Scan ID	24	37	40	55	63	65	69	83	97	105	106	110	114	118	122	Average
PGSR	0.38	0.67	0.42	0.35	0.76	0.58	0.53	1.08	0.70	0.61	0.45	0.55	0.34	0.42	0.40	0.55
QGS	0.45	0.77	0.37	0.37	0.84	0.91	0.55	1.15	1.00	0.62	0.46	0.70	0.36	0.45	0.41	0.63
GOF	0.53	0.89	0.43	0.38	1.33	0.87	0.77	1.28	1.29	0.79	0.77	1.15	0.46	0.70	0.54	0.81
SOF	0.55	0.70	0.41	0.39	1.12	0.73	0.66	1.11	1.47	0.60	0.63	1.05	0.58	0.62	0.48	0.74
MILo	0.46	0.77	0.35	0.39	0.80	0.86	0.72	1.20	1.22	0.66	0.63	0.93	0.38	0.79	0.49	0.71
Ours	0.43	0.70	0.34	0.39	0.92	0.70	0.67	1.03	1.19	0.50	0.49	1.05	0.47	0.41	0.40	0.65

Appendix

Table 7: Full Geometry Reconstruction results for the DTU dataset [ 23 ] , where we report the Chamfer distance (lower is better). Our method attains the lowest average Chamfer distance for all methods designed to extract unbounded meshes.

	Mip-NeRF 360 Indoor			Mip-NeRF 360 Outdoor
	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	# Gaussians
PGSR $\dagger$	30.35	0.924	0.176	24.29	0.718	0.236	-
QGS $\dagger$	30.45	0.919	0.184	24.32	0.706	0.242	-
GOF $\dagger$	30.47	0.918	0.189	24.79	0.745	0.208	-
SOF $\dagger$	30.13	0.916	0.188	24.78	0.745	0.208	-
MILo $\dagger$	29.96	0.920	0.191	24.47	0.718	0.290	-

Appendix

Table 8: Novel View Synthesis Ablation for the Mip-NeRF 360 dataset [ 1 ] , where we report PSNR, SSIM and LPIPS. Notably, adding

{\mathcal{L}}_{\text{conf}}

does not negatively impact image quality metrics overall. Results with

\dagger

were taken from [ 77 , 52 , 18 ] .

Panel of Figure 12 — Figure 12: Confidence Evaluation: Our learned confidence directly correlates with the error maps, demonstrating that it can detect under-reconstructed regions.

Panel of Figure 13 — Figure 13: Effect of our proposed Variance Losses: We visualize the first-hit , last-hit and fully alpha-blended colors and normals. As can be seen, our variance losses constrain spurious geometry and lead to smoother normal maps.

Panel of Figure 14 — Figure 14: Effect of our SSIM-decoupled Appearance Module : We show rendered images from two spatially close camera poses. As can be seen, the illumination for Ours is more consistent; this is caused by our improved appearance module.

	Tanks & Temples				Mip-NeRF 360
	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	# Verts $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	# Verts $\downarrow$
Ours	21.186	0.683	0.345	13.06M	23.639	0.661	0.332	18.43M
MILo	21.169	0.673	0.354	5.34M	23.869	0.675	0.322	6.20M
SOF	20.936	0.672	0.354	14.07M	23.304	0.634	0.351	23.28M
GOF	20.959	0.675	0.354	14.40M	22.998	0.619	0.365	22.37M

Appendix

Table 9: Mesh-Based Novel View Synthesis: We report novel view synthesis metrics as a proxy for evaluating underlying geometric accuracy, using the Mip-NeRF 360 and Tanks & Temples datasets. We attain comparable quality to MILo [ 18 ] across both tested datasets.

3D Mesh Reconstruction

3D mesh reconstruction turns images, depth, or sensor data into explicit surfaces for measurement, editing, and simulation.

See all 109 papers

MEGA: Object-Level Mesh Extraction from 3D Gaussian Splatting via Spatial Visual Distillation
Oct 1, 2026Liwei Liao, Yingkui Zhang, Qianqian Tong +13D Gaussian Splatting 3D Reconstruction
Peking University · Pengcheng Laboratory · Beihang University
Mesh2GS: White-Box 3DGS Construction via Plenoptic Sampling
Jun 20, 2026Haoran Zhu, Youcheng Cai, Huangsheng Du +23D Gaussian Splatting Language-Embedded 3DGS
University of Science and Technology of China
From Blobs to Spokes: High-Fidelity Surface Reconstruction via Oriented Gaussians
Apr 8, 2026Diego Gomez, Antoine Guédon, Nissim Maruani +23D Gaussian Splatting 3D Mesh Reconstruction
LIX, École Polytechnique, France · Inria, Côte d’Azur, France

Oct 1, 2026J/K move · Enter open · S save

MEGA: Object-Level Mesh Extraction from 3D Gaussian Splatting via Spatial Visual Distillation

Liwei Liao, Yingkui Zhang, Qianqian Tong, Ronggang Wang

Peking University · Pengcheng Laboratory · Beihang University

3D Gaussian Splatting 3D Reconstruction 3D Mesh Reconstruction

Open paper AskPDF arXiv

Mesh extraction from 3D Gaussian Splatting (3DGS) aims to endow 3D Gaussians with accurate geometric structures, enabling explicit and precise 3D occupancy. However, existing methods primarily focus on scene-level mesh extraction, making them unable to represent object-level occupancy and often resulting in non-watertight surfaces. To overcome these limitations, we propose MEGA (Mesh Extraction from GAussians), a ``segment-then-mesh'' framework for extracting object-level, watertight meshes from complex 3DGS scenes. At the core of MEGA are Spatial Visual Distillation (SVD) and a mask-guided neural surface reconstruction module. SVD treats the 3DGS model as a teacher, sampling diverse camera poses and rendering the corresponding views of each segmented object. These observations are then used to train a mesh reconstruction model through photometric supervision. Extensive experiments on several widely used benchmarks demonstrate that MEGA achieves state-of-the-art performance in recovering accurate object-level 3D occupancy. Moreover, MEGA enables complex physical interactions by combining high-quality object-level meshes for geometric occupancy with 3DGS representations for photorealistic rendering.

Abstract

3D Mesh Reconstruction

Explore similar work

MEGA: Object-Level Mesh Extraction from 3D Gaussian Splatting via Spatial Visual Distillation

Mesh2GS: White-Box 3DGS Construction via Plenoptic Sampling

From Blobs to Spokes: High-Fidelity Surface Reconstruction via Oriented Gaussians