Confidence-Based Mesh Extraction from 3D Gaussians

Lukas Radl*1   Felix Windisch*1   Andreas Kurz*1  
Thomas Köhler 1   Michael Steiner 1   Markus Steinberger 1,2
* denotes equal contribution!
1 Graz University of Technology   2 Huawei Technologies  
1 2
Paper Video Code (coming soon) Data (coming soon) Meshes (coming soon)

Teaser: We present a novel method for fast, unbounded surface extraction. Our meshes are detailed, artifact-free, and can be extracted in roughly 20 minutes!

Self-Supervised Confidence Learning

We introduce a self-supervised confidence framework to 3DGS, allowing the model to render per-pixel confidences. This automatically assigns low confidence (red) to uncertain areas, e.g. reflections.
RGB Confidence
RGB Confidence
RGB Confidence
RGB Confidence
RGB Confidence
RGB Confidence

Variance Losses

We additionally propose variance losses, which constrains individual Gaussians, removing spurious geometry and increasing normal smoothness. Here, we visualize "first-hit" Gaussians, i.e. the color/normal of the first blended primitive.
w/ Variance Losses w/o Variance Losses
w/ Variance Losses w/o Variance Losses

SSIM-Decoupled Apperance

We additionally propose an SSIM-decoupled appearance module, which allows our model to better disentangle camera-dependent exposure and geometry. For VastGaussian [Lin et al. 2024], the lighting variation is baked into the 3DGS point cloud, and manifests as view-dependent appearance.
Visualization Details: We pick two nearby views from the same scene, and perform a dense warp using RoMa [Edstedt et al. 2024]. Inspect the varying appearance for VastGaussian using this image slider.

Ours (SSIM-Decoupled)

Original View Ours - Warped View

VastGaussian

Original View VastGaussian - Warped View

Detailed Unbounded Mesh Extraction

Compared to the state-of-the-art unbounded mesh extraction method MILo [Guedon et al. 2025], our meshes exhibit significantly finer details. Note that ours pipeline is also 3x faster than MILo.
Ours MILo
Ours MILo
Ours MILo
Ours MILo
Ours MILo
Ours MILo

Comparison with Bounded Methods

Compared to PGSR [Chen et al. 2024], a representative multi-view bounded method, our meshes are more complete, and exhibit vastly finer details.
Ours PGSR
Ours PGSR
Ours PGSR
Ours PGSR
Ours PGSR
Ours PGSR

Video Comparisons

Compare our method (left) against either MILo [Guedon et al. 2025] or PGSR [Chen et al. 2024] (right). Drag the vertical divider in each scene.
Comparison method
Scene

Barn

Related Work

[Guédon et al. 2025]

MILo: Mesh-In-the-Loop Gaussian Splatting for Detailed and Efficient Surface Reconstruction. ACM TOG 44(6). 2025.

[Radl et al. 2025]

SOF: Sorted Opacity Fields for Fast Unbounded Surface Reconstruction. SIGGRAPH Asia. 2025.

[Chen et al. 2024]

PGSR: Planar-based Gaussian Splatting for Efficient and High-Fidelity Surface Reconstruction. TVCG 31(9). 2024.

[Lin et al. 2024]

VastGaussian: Vast 3D Gaussians for Large Scene Reconstruction. CVPR. 2024.

BibTeX

If you find our work useful, consider using a citation (use the copy & paste button in the top right corner): 

@misc{radl2026come,
  author        = {Radl, Lukas and Windisch, Felix and Kurz, Andreas and K{\"o}hler, Thomas and Steiner, Michael and Steinberger, Markus},
  title         = {{Confidence-Based Mesh Extraction from 3D Gaussians}},
  year          = {2026},
  eprint        = {2603.24725},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2603.24725}, 
}