pith. sign in

arxiv: 2604.07337 · v1 · submitted 2026-04-08 · 💻 cs.CV

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

Diego Gomez , Antoine Gu\'edon , Nissim Maruani , Bingchen Gong , Maks Ovsjanikov This is my paper

Pith reviewed 2026-05-10 17:50 UTC · model grok-4.3

classification 💻 cs.CV
keywords 3D Gaussian SplattingSurface ReconstructionOriented NormalsWatertight MeshesOccupancy FieldThin Structure RecoveryNovel View SynthesisDensification Strategy
0
0 comments X

The pith

Oriented Gaussians with learnable normals produce complete watertight meshes from 3D splatting

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper seeks to fix the core limitation of 3D Gaussian Splatting, where opacity blending prevents reliable surface extraction and forces heuristic depth fusion. It derives a continuous occupancy field from each Gaussian by adding a learnable oriented normal and an adapted attenuation function, then enforces surface wrapping through a consistency loss and targeted densification. This produces closed, high-detail meshes including thin elements that prior methods miss. A reader would care because the approach keeps the rendering speed of splatting while delivering the geometric completeness normally associated with slower implicit representations.

Core claim

Gaussian Wrapping introduces a learnable oriented normal per Gaussian and an adapted attenuation model that yields closed-form expressions for both the normal field and an occupancy field at any point in space. A consistency loss plus a dedicated densification strategy force every Gaussian to lie on and enclose the true surface, closing holes and eliminating spurious geometry. The modified rasterizer outputs depth as an isosurface of this field, and Primal Adaptive Meshing extracts watertight meshes at chosen resolutions; the method reports new state-of-the-art results on DTU and Tanks and Temples while recovering thin structures such as bicycle spokes using far fewer primitives than prior S

What carries the argument

The learnable oriented normal together with the adapted attenuation formulation, which supplies closed-form occupancy and normal fields used to define an isosurface for mesh extraction

If this is right

  • Standard surface evaluation protocols contain measurable biases that the proposed alternatives avoid
  • Meshes of arbitrary resolution can be extracted on demand for any region of interest via primal adaptive meshing
  • Complete watertight meshes are obtained at a small fraction of the primitive count required by concurrent methods
  • Thin structures that defeat existing approaches become recoverable as part of a single closed surface

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same oriented-Gaussian occupancy field could be queried directly for ray-tracing or collision queries without an intermediate mesh
  • Because the representation remains compact, it may scale to city-scale scenes where memory limits currently force downsampling
  • The formulation might transfer to dynamic or time-varying splatting by adding a temporal consistency term on the oriented normals

Load-bearing premise

That learnable oriented normals, adapted attenuation, a consistency loss, and densification will place every Gaussian exactly on the true surface and produce a closed shell without holes or extra geometry

What would settle it

Reconstruct the bicycle scene from the DTU dataset and check whether the extracted mesh still contains holes or missing spokes when measured with the two rigorous evaluation protocols introduced in the paper

Figures

Figures reproduced from arXiv: 2604.07337 by Antoine Gu\'edon, Bingchen Gong, Diego Gomez, Maks Ovsjanikov, Nissim Maruani.

Figure 1
Figure 1. Figure 1: High-fidelity surface reconstruction from 3D Gaussian Splatting. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Examples of textured meshes reconstructed with our method. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Wrapping the surface with Oriented Gaussians. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Primal Adaptive Meshing on restricted scene segments. [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison on T&T and MipNerf360. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison between our continuous attenuation model, and the one [PITH_FULL_IMAGE:figures/full_fig_p022_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Equivalence between Gaussian Splatting rendering and our [PITH_FULL_IMAGE:figures/full_fig_p025_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Evolution of the vacancy along a ray. In contrast to the initial attenuation we derived for Gaussian Splatting (a), our oriented Gaussian formulation (b) allows for a reciprocal formulation of the vacancy. This figure shows how our formulation corrects the bias present in (a). Indeed, in (a) vacancy starts changing just before each Gaussian. This poses a problem for G2, since vacancy starts evolving from o… view at source ↗
Figure 9
Figure 9. Figure 9: Gaussian vector field aligned with surface normals. [PITH_FULL_IMAGE:figures/full_fig_p031_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Effect of the Newton projection step on vacancy values at sampled [PITH_FULL_IMAGE:figures/full_fig_p035_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative mesh renders on DTU. Mesh renders of our method on three DTU scenes. GT Point Cloud PGSR Ours [PITH_FULL_IMAGE:figures/full_fig_p038_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: T&T evaluation shortcomings on the Caterpillar scene. [PITH_FULL_IMAGE:figures/full_fig_p038_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Evaluation Code of T&T used by the literature [10, 23, 55, 56] [PITH_FULL_IMAGE:figures/full_fig_p039_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Ablation of the normal field loss LN and densification on the Barn and Truck scenes. On the Barn scene, we see that by extracting a mesh with 2 pivots without our alignment loss LN , several artifacts are present. These come in the form of large triangles that stem from a bad orientation of the Gaussians which result in bad placement for the pivots. Then below, we ablate the need for our densification (Ou… view at source ↗
Figure 15
Figure 15. Figure 15: Qualitative effect of Normal Field (NF) supervision on RaDe-GS. [PITH_FULL_IMAGE:figures/full_fig_p042_15.png] view at source ↗
read the original abstract

3D Gaussian Splatting (3DGS) has revolutionized fast novel view synthesis, yet its opacity-based formulation makes surface extraction fundamentally difficult. Unlike implicit methods built on Signed Distance Fields or occupancy, 3DGS lacks a global geometric field, forcing existing approaches to resort to heuristics such as TSDF fusion of blended depth maps. Inspired by the Objects as Volumes framework, we derive a principled occupancy field for Gaussian Splatting and show how it can be used to extract highly accurate watertight meshes of complex scenes. Our key contribution is to introduce a learnable oriented normal at each Gaussian element and to define an adapted attenuation formulation, which leads to closed-form expressions for both the normal and occupancy fields at arbitrary locations in space. We further introduce a novel consistency loss and a dedicated densification strategy to enforce Gaussians to wrap the entire surface by closing geometric holes, ensuring a complete shell of oriented primitives. We modify the differentiable rasterizer to output depth as an isosurface of our continuous model, and introduce Primal Adaptive Meshing for Region-of-Interest meshing at arbitrary resolution. We additionally expose fundamental biases in standard surface evaluation protocols and propose two more rigorous alternatives. Overall, our method Gaussian Wrapping sets a new state-of-the-art on DTU and Tanks and Temples, producing complete, watertight meshes at a fraction of the size of concurrent work-recovering thin structures such as the notoriously elusive bicycle spokes.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper introduces Gaussian Wrapping, an extension of 3D Gaussian Splatting for surface reconstruction. It adds learnable oriented normals per Gaussian and an adapted attenuation to derive closed-form expressions for occupancy and normal fields. A novel consistency loss and dedicated densification strategy are proposed to force Gaussians to wrap the full surface and close holes, enabling watertight mesh extraction via a modified rasterizer and Primal Adaptive Meshing. The method claims SOTA quantitative results on DTU and Tanks & Temples, superior recovery of thin structures (e.g., bicycle spokes), and exposes biases in standard surface evaluation protocols.

Significance. If the derivations hold and the empirical results are robust, the work would be significant for bridging explicit Gaussian representations with accurate, complete surface extraction. It offers a more principled alternative to TSDF fusion heuristics while retaining fast rendering, with potential impact on real-time 3D applications. The new evaluation protocols and thin-structure recovery are notable strengths if substantiated.

major comments (3)
  1. [Abstract, §4] Abstract and §4 (method): The central claim that the combination of oriented normals, adapted attenuation, consistency loss, and densification produces complete, hole-free watertight meshes rests on the assumption that these components force every Gaussian onto the true surface. No derivation or convergence analysis is provided showing that the loss landscape avoids local minima with gaps or spurious geometry; this is load-bearing for the bicycle-spokes result and the 'complete shell' guarantee.
  2. [Results] Results section (DTU/Tanks & Temples tables): The SOTA claims and thin-structure recovery are reported quantitatively, but without the full loss formulations, ablation tables on the consistency loss and densification hyperparameters, or failure-case analysis, it is impossible to assess whether the improvements are due to the closed-form fields or post-hoc tuning. This directly affects soundness of the 'parameter-free' and 'principled' framing.
  3. [§3] §3 (occupancy derivation): While closed-form expressions for occupancy and normals are derived from the oriented primitives, the occupancy field is then regularized via the consistency loss rather than being directly fitted; this introduces circularity between the model and the training objective that is not analyzed for stability on thin or complex geometry.
minor comments (2)
  1. [§3] Notation for the adapted attenuation and oriented normal parameters should be introduced with explicit equations early in §3 to improve readability.
  2. [Method] The Primal Adaptive Meshing procedure is described at a high level; a pseudocode or step-by-step algorithm would clarify how it differs from standard marching cubes.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major comment point-by-point below, providing clarifications on our methodological choices and empirical support while committing to revisions that strengthen the presentation without overstating theoretical guarantees.

read point-by-point responses

  1. Referee: [Abstract, §4] Abstract and §4 (method): The central claim that the combination of oriented normals, adapted attenuation, consistency loss, and densification produces complete, hole-free watertight meshes rests on the assumption that these components force every Gaussian onto the true surface. No derivation or convergence analysis is provided showing that the loss landscape avoids local minima with gaps or spurious geometry; this is load-bearing for the bicycle-spokes result and the 'complete shell' guarantee.

    Authors: We agree that no formal convergence analysis or derivation of the loss landscape is provided, and this is a genuine limitation when claiming a 'complete shell' guarantee. The consistency loss penalizes occupancy inconsistencies across views while the densification strategy explicitly targets regions with insufficient Gaussian coverage to close holes. We will revise §4 to include additional discussion of the loss design rationale, supported by new visualizations of intermediate optimization states on thin structures such as bicycle spokes. We will also moderate language in the abstract and method sections to avoid implying a theoretical guarantee. revision: partial

  2. Referee: [Results] Results section (DTU/Tanks & Temples tables): The SOTA claims and thin-structure recovery are reported quantitatively, but without the full loss formulations, ablation tables on the consistency loss and densification hyperparameters, or failure-case analysis, it is impossible to assess whether the improvements are due to the closed-form fields or post-hoc tuning. This directly affects soundness of the 'parameter-free' and 'principled' framing.

    Authors: The loss formulations are already detailed in §4.2 (consistency loss) and §4.3 (densification). To address the concern, we will add comprehensive ablation tables in the supplementary material examining the consistency loss weight and densification hyperparameters. We will also include a dedicated failure-case analysis section with quantitative and qualitative examples to demonstrate that gains stem from the closed-form fields rather than extensive tuning, thereby supporting the principled framing. revision: yes

  3. Referee: [§3] §3 (occupancy derivation): While closed-form expressions for occupancy and normals are derived from the oriented primitives, the occupancy field is then regularized via the consistency loss rather than being directly fitted; this introduces circularity between the model and the training objective that is not analyzed for stability on thin or complex geometry.

    Authors: The closed-form occupancy and normal expressions in §3 are derived solely from the oriented Gaussian primitives and adapted attenuation, prior to and independent of the loss. The consistency loss then enforces multi-view agreement between this derived field and observed surfaces, which is a standard regularization step rather than circular. We will add a new paragraph in the revised §3 analyzing stability on thin and complex geometry, including empirical metrics from our DTU and Tanks & Temples experiments showing no observed instability. revision: partial

Circularity Check

0 steps flagged

No circularity: forward derivation of occupancy from oriented primitives plus independent training components

full rationale

The paper presents a direct mathematical derivation of closed-form occupancy and normal fields from the combination of learnable per-Gaussian oriented normals and an adapted attenuation function. This is a constructive step (new primitives → explicit field expressions) rather than any reduction of the output to the input by definition or fitting. The subsequent consistency loss and densification strategy are explicitly introduced as additional optimization mechanisms to encourage surface wrapping; they are not claimed to be derived from the occupancy equations themselves. No self-citations, uniqueness theorems, or renamed empirical patterns are invoked as load-bearing justifications for the core field expressions. The overall pipeline therefore remains self-contained against external benchmarks and does not exhibit any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 1 invented entities

The central claim rests on a new per-Gaussian oriented normal (learnable), an adapted attenuation model whose closed-form properties are asserted, and empirical densification rules whose justification is not supplied in the abstract.

free parameters (2)
  • learnable oriented normals
    Direction vector attached to each Gaussian and optimized during training; directly affects the derived occupancy field.
  • densification and consistency-loss hyperparameters
    Control how aggressively Gaussians are added or moved to close geometric holes.
axioms (1)
  • domain assumption Adapted attenuation formulation yields closed-form normal and occupancy fields at arbitrary points
    Invoked to justify the continuous geometric field without numerical integration.
invented entities (1)
  • oriented Gaussian primitive no independent evidence
    purpose: To supply both color/opacity and a local surface normal for the occupancy derivation
    New primitive type introduced; no independent evidence supplied beyond empirical mesh quality.

pith-pipeline@v0.9.0 · 5576 in / 1380 out tokens · 30283 ms · 2026-05-10T17:50:17.407993+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

65 extracted references · 65 canonical work pages

  1. [1]

    In: Proceedings of the 25th annual conference on Computer graph- ics and interactive techniques - SIGGRAPH ’98

    Amenta, N., Bern, M., Kamvysselis, M.: A new Voronoi-based surface reconstruc- tion algorithm. In: Proceedings of the 25th annual conference on Computer graph- ics and interactive techniques - SIGGRAPH ’98. pp. 415–421. ACM Press (1998)

    work page 1998

  2. [2]

    In: Proceedings of the sixth ACM symposium on Solid modeling and applications

    Amenta, N., Choi, S., Kolluri, R.K.: The power crust. In: Proceedings of the sixth ACM symposium on Solid modeling and applications. pp. 249–266. SMA ’01, As- sociation for Computing Machinery (2001)

    work page 2001

  3. [3]

    ACM computing surveys (CSUR)23(3), 345–405 (1991)

    Aurenhammer, F.: Voronoi diagrams—a survey of a fundamental geometric data structure. ACM computing surveys (CSUR)23(3), 345–405 (1991)

    work page 1991

  4. [4]

    Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin Brualla, R., Srini- vasan, P.P.: Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Ra- diance Fields. In: Int. Conf. Comput. Vis. pp. 5855–5864 (2021)

    work page 2021

  5. [5]

    In: IEEE Conf

    Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5460–5469. IEEE (2022)

    work page 2022

  6. [6]

    ACM Trans

    Binninger, A., Wiersma, R., Herholz, P., Sorkine-Hornung, O.: TetWeave: Isosur- face Extraction using On-The-Fly Delaunay Tetrahedral Grids for Gradient-Based Mesh Optimization. ACM Trans. Graph.44(4), 1–19 (2025)

    work page 2025

  7. [7]

    In: 2025 IEEE/CVF Winter Con- ference on Applications of Computer Vision (WACV)

    Blanc, H., Deschaud, J.E., Paljic, A.: Raygauss: Volumetric Gaussian-based Ray Casting for Photorealistic Novel View Synthesis. In: 2025 IEEE/CVF Winter Con- ference on Applications of Computer Vision (WACV). pp. 1808–1817. IEEE (2025)

    work page 2025

  8. [8]

    Celarek, A., Kopanas, G., Drettakis, G., Wimmer, M., Kerbl, B.: Does 3d gaus- sian splatting need accurate volumetric rendering? In: Computer Graphics Forum. vol. 44, p. e70032. Wiley Online Library (2025)

    work page 2025

  9. [9]

    Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: TensoRF: Tensorial Radiance Fields. In: Eur. Conf. Comput. Vis. pp. 333–350 (2022)

    work page 2022

  10. [10]

    Chen, D., Li, H., Ye, W., Wang, Y., Xie, W., Zhai, S., Wang, N., Liu, H., Bao, H., Zhang, G.: PGSR: Planar-based Gaussian Splatting for Efficient and High-Fidelity SurfaceReconstruction.IEEETransactionsonVisualizationandComputerGraph- ics31(9), 6100–6111 (2024)

    work page 2024

  11. [11]

    Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance, 2025

    Chen, H., Li, C., Lee, G.H.: NeuSG: Neural Implicit Surface Reconstruction with 3D Gaussian Splatting Guidance. arXiv preprint arXiv:2312.00846 (2023)

    work page arXiv 2023

  12. [12]

    Chen, H., Wei, F., Li, C., Huang, T., Wang, Y., Lee, G.H.: VCR-GauS: View Consistent Depth-Normal Regularizer for Gaussian Surface Reconstruction. In: Adv. Neural Inform. Process. Syst. (2024)

    work page 2024

  13. [13]

    In: Proceedings of the eighth ACM symposium on Solid modeling and applications

    Dey, T.K., Goswami, S.: Tight cocone: a water-tight surface reconstructor. In: Proceedings of the eighth ACM symposium on Solid modeling and applications. pp. 127–134. ACM (2003)

    work page 2003

  14. [14]

    The Visual Computer 24(6), 411–422 (2008)

    Dey, T.K., Levine, J.A.: Delaunay meshing of isosurfaces. The Visual Computer 24(6), 411–422 (2008)

    work page 2008

  15. [15]

    In: IEEE Conf

    Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenox- els: Radiance Fields without Neural Networks. In: IEEE Conf. Comput. Vis. Pat- tern Recog. pp. 5501–5510 (2022)

    work page 2022

  16. [16]

    In: Computer Graphics International Conference

    Gomez, D., Philip, J., Kaiser, A., Michel, É.: RRM: Relightable assets using Radi- ance guided Material extraction. In: Computer Graphics International Conference. pp. 17–41. Springer (2024)

    work page 2024

  17. [17]

    Govindarajan, S., Rebain, D., Yi, K.M., Tagliasacchi, A.: Radiant Foam: Real- Time Differentiable Ray Tracing. In: Int. Conf. Comput. Vis. (2025) 18 D. Gomez et al

    work page 2025

  18. [18]

    ACM Trans

    Guédon, A., Gomez, D., Maruani, N., Gong, B., Drettakis, G., Ovsjanikov, M.: MILo: Mesh-In-the-Loop Gaussian Splatting for Detailed and Efficient Surface Reconstruction. ACM Trans. Graph.44(6), 1–15 (2025)

    work page 2025

  19. [19]

    In: IEEE Conf

    Guédon, A., Lepetit, V.: SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering. In: IEEE Conf. Com- put. Vis. Pattern Recog. pp. 5354–5363 (2024)

    work page 2024

  20. [20]

    arXiv preprint arXiv:2512.06818 (2025)

    Held, J., Son, S., Vandeghen, R., Rebain, D., Gadelha, M., Zhou, Y., Cioppa, A., Lin, M.C., Van Droogenbroeck, M., Tagliasacchi, A.: MeshSplatting: Differentiable Rendering with Opaque Meshes. arXiv preprint arXiv:2512.06818 (2025)

    work page arXiv 2025

  21. [21]

    In: ACM SIGGRAPH 2024 conference pa- pers

    Huang, B., Yu, Z., Chen, A., Geiger, A., Gao, S.: 2D Gaussian Splatting for Ge- ometrically Accurate Radiance Fields. In: ACM SIGGRAPH 2024 conference pa- pers. pp. 1–11 (2024)

    work page 2024

  22. [22]

    In: IEEE Conf

    Jensen, R., Dahl, A., Vogiatzis, G., Tola, E., Aanæs, H.: Large scale multi-view stereopsis evaluation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 406–413 (2014)

    work page 2014

  23. [23]

    In: IEEE Conf

    Jiang, K., Sivaram, V., Peng, C., Ramamoorthi, R.: Geometry Field Splatting with Gaussian Surfels. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5752– 5762 (2025)

    work page 2025

  24. [24]

    In: IEEE Conf

    Jin, H., Liu, I., Xu, P., Zhang, X., Han, S., Bi, S., Zhou, X., Xu, Z., Su, H.: TensoIR: Tensorial Inverse Rendering. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 165– 174 (2023)

    work page 2023

  25. [25]

    ACM Trans

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Trans. Graph.42(4) (July 2023)

    work page 2023

  26. [26]

    Kheradmand, S., Rebain, D., Sharma, G., Sun, W., Tseng, Y.C., Isack, H., Kar, A., Tagliasacchi, A., Yi, K.M.: 3D gaussian splatting as markov chain monte carlo. Adv. Neural Inform. Process. Syst.37, 80965–80986 (2024)

    work page 2024

  27. [27]

    ACM Transactions on Graphics36(4) (2017)

    Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics36(4) (2017)

    work page 2017

  28. [28]

    In: IEEE Conf

    Li, S., Liu, Y.S., Han, Z.: Gaussianudf: Inferring unsigned distance functions through 3d gaussian splatting. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 27113–27123 (2025)

    work page 2025

  29. [29]

    In: IEEE Conf

    Li, Z., Müller, T., Evans, A., Taylor, R.H., Unberath, M., Liu, M.Y., Lin, C.H.: Neuralangelo: High-fidelity neural surface reconstruction. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 8456–8465 (2023)

    work page 2023

  30. [30]

    In: IEEE Conf

    Liang, Z., Zhang, Q., Feng, Y., Shan, Y., Jia, K.: GS-IR: 3D Gaussian Splatting for Inverse Rendering. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 21644–21653 (2024)

    work page 2024

  31. [31]

    ACM Trans

    Lyu, X., Sun, Y.T., Huang, Y.H., Wu, X., Yang, Z., Chen, Y., Pang, J., Qi, X.: 3DGSR: Implicit Surface Reconstruction with 3D Gaussian Splatting. ACM Trans. Graph.43(6), 1–12 (2024)

    work page 2024

  32. [32]

    Mai, A., Hedman, P., Kopanas, G., Verbin, D., Futschik, D., Xu, Q., Kuester, F., Barron, J.T., Zhang, Y.: Ever: Exact volumetric ellipsoid rendering for real-time view synthesis. In: Int. Conf. Comput. Vis. pp. 4930–4939 (2025)

    work page 2025

  33. [33]

    Mai, A., Hedstrom, T., Kopanas, G., Kontkanen, J., Kuester, F., Barron, J.T.: Radiance Meshes for Volumetric Reconstruction (2025)

    work page 2025

  34. [34]

    Mai, A., Verbin, D., Kuester, F., Fridovich-Keil, S.: Neural Microfacet Fields for Inverse Rendering (2023)

    work page 2023

  35. [35]

    In: SIG- GRAPH Asia Conference Proceedings

    Mallick, S.S., Goel, R., Kerbl, B., Steinberger, M., Carrasco, F.V., De La Torre, F.: Taming 3DGS: High-quality Radiance Fields with Limited Resources. In: SIG- GRAPH Asia Conference Proceedings. pp. 1–11 (2024) Gaussian Wrapping 19

    work page 2024

  36. [36]

    Maruani, N., Klokov, R., Ovsjanikov, M., Alliez, P., Desbrun, M.: VoroMesh: Learning Watertight Surface Meshes with Voronoi Diagrams. In: Int. Conf. Com- put. Vis. pp. 14565–14574 (2023)

    work page 2023

  37. [37]

    In: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Maruani,N.,Ovsjanikov,M.,Alliez,P.,Desbrun,M.:PoNQ:ANeuralQEM-Based Mesh Representation. In: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3647–3657 (Jun 2024)

    work page 2024

  38. [38]

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In: Eur. Conf. Comput. Vis. pp. 99–106 (2020)

    work page 2020

  39. [39]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Miller, B., Chen, H., Lai, A., Gkioulekas, I.: Objects as Volumes: A Stochastic Geometry View of Opaque Solids. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 87–97 (June 2024)

    work page 2024

  40. [40]

    ACM Trans

    Müller, T., Evans, A., Schied, C., Keller, A.: Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM Trans. Graph.41(4), 102:1–102:15 (Jul 2022)

    work page 2022

  41. [41]

    In: IEEE Conf

    Munkberg, J., Hasselgren, J., Shen, T., Gao, J., Chen, W., Evans, A., Müller, T., Fidler, S.: Extracting Triangular 3D Models, Materials, and Lighting From Images. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)

    work page 2022

  42. [42]

    In: SIGGRAPH Asia Conference Proceedings (2025)

    Radl, L., Windisch, F., Deixelberger, T., Hladky, J., Steiner, M., Schmalstieg, D., Steinberger, M.: SOF: Sorted Opacity Fields for Fast Unbounded Surface Rcon- struction. In: SIGGRAPH Asia Conference Proceedings (2025)

    work page 2025

  43. [43]

    Reiser, C., Garbin, S., Srinivasan, P.P., Verbin, D., Szeliski, R., Mildenhall, B., Barron, J.T., Hedman, P., Geiger, A.: Binary Opacity Grids: Capturing Fine Geo- metricDetailforMesh-BasedViewSynthesis.SIGGRAPHConferenceProceedings (2024)

    work page 2024

  44. [44]

    arXiv preprint arXiv:2510.16837 (2025)

    Ren, H., Yan, Q., Lu, M., Lu, R., Zhu, Z.: 2DGS-R: Revisiting the Normal Consis- tency Regularization in 2D Gaussian Splatting. arXiv preprint arXiv:2510.16837 (2025)

    work page arXiv 2025

  45. [45]

    ACM Trans

    Shen, T., Munkberg, J., Hasselgren, J., Yin, K., Wang, Z., Chen, W., Gojcic, Z., Fidler, S., Sharp, N., Gao, J.: Flexible Isosurface Extraction for Gradient-Based Mesh Optimization. ACM Trans. Graph.42(4) (2023)

    work page 2023

  46. [46]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Son, S., Gadelha, M., Zhou, Y., Fisher, M., Xu, Z., Qiao, Y.L., Lin, M.C., Zhou, Y.: DMesh++: An Efficient Differentiable Mesh for Complex Shapes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 26590–26599 (2025)

    work page 2025

  47. [47]

    Advances in Neural Information Processing Systems37, 12035–12077 (2024)

    Son, S., Gadelha, M., Zhou, Y., Xu, Z., Lin, M., Zhou, Y.: DMesh: A Differentiable Mesh Representation. Advances in Neural Information Processing Systems37, 12035–12077 (2024)

    work page 2024

  48. [48]

    In: Computer Graphics Forum

    Sulzer, R., Landrieu, L., Marlet, R., Vallet, B.: Scalable Surface Reconstruction with Delaunay-Graph Neural Networks. In: Computer Graphics Forum. vol. 40, pp. 157–167. Wiley Online Library (2021)

    work page 2021

  49. [49]

    IEEE Conf

    Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-NeRF: Structured View-Dependent Appearance for Neural Radiance Fields. IEEE Conf. Comput. Vis. Pattern Recog. (2022)

    work page 2022

  50. [50]

    Neural Inform

    Wang, P., Liu, L., Liu, Y., Theobalt, C., Komura, T., Wang, W.: NeuS: Learning neuralimplicitsurfacesbyvolumerenderingformulti-viewreconstruction.In:Adv. Neural Inform. Process. Syst. (2021)

    work page 2021

  51. [51]

    Yariv, L., Gu, J., Kasten, Y., Lipman, Y.: Volume rendering of neural implicit surfaces. In: Adv. Neural Inform. Process. Syst. (2021) 20 D. Gomez et al

    work page 2021

  52. [52]

    In: ACM SIGGRAPH 2023 conference proceedings

    Yariv, L., Hedman, P., Reiser, C., Verbin, D., Srinivasan, P.P., Szeliski, R., Bar- ron, J.T., Mildenhall, B.: BakedSDF: Meshing Neural SDFs for Real-Time View Synthesis. In: ACM SIGGRAPH 2023 conference proceedings. pp. 1–9 (2023)

    work page 2023

  53. [53]

    Yu, M., Lu, T., Xu, L., Jiang, L., Xiangli, Y., Dai, B.: GSDF: 3DGS Meets SDF for Improved Rendering and Reconstruction. In: Adv. Neural Inform. Process. Syst. (2024)

    work page 2024

  54. [54]

    In: IEEE Conf

    Yu, Z., Chen, A., Huang, B., Sattler, T., Geiger, A.: Mip-Splatting: Alias-free 3D Gaussian Splatting. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 19447–19456 (2024)

    work page 2024

  55. [55]

    ACM Transactions on Graphics (ToG)43(6), 1–13 (2024)

    Yu, Z., Sattler, T., Geiger, A.: Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics (ToG)43(6), 1–13 (2024)

    work page 2024

  56. [56]

    ACM Trans

    Zhang, B., Fang, C., Shrestha, R., Liang, Y., Long, X., Tan, P.: RaDe-GS: Raster- izing Depth in Gaussian Splatting. ACM Trans. Graph. (2024)

    work page 2024

  57. [57]

    arXiv preprint arXiv:2601.17835 (2026)

    Zhang, B., Jiang, C., Li, H., Shen, S., Tan, P.: Geometry-Grounded Gaussian Splatting. arXiv preprint arXiv:2601.17835 (2026)

    work page arXiv 2026

  58. [58]

    In: 2025 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR)

    Zhang, C., Wang, W., Li, X., Liao, X., Su, W., Tao, W.: High-Fidelity Lightweight Mesh Reconstruction from Point Clouds. In: 2025 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR). pp. 11739–11748 (Jun 2025), iSSN: 2575-7075

    work page 2025

  59. [59]

    Zhang, W., Liu, Y.S., Han, Z.: Neural Signed Distance Function Inference through Splatting 3D Gaussians Pulled on Zero-Level Set. In: Adv. Neural Inform. Process. Syst. (2024)

    work page 2024

  60. [60]

    Zhang, Z., Huang, B., Jiang, H., Zhou, L., Xiang, X., Shen, S.: Quadratic Gaussian Splatting for Efficient and Detailed Surface Reconstruction. In: Int. Conf. Comput. Vis. (2025) Gaussian Wrapping 21 A Overview of the Appendix This appendix serves three purposes:(i)provide proofs for the theoretical claims made in the main paper;(ii)provide the implementa...

    work page 2025

  61. [61]

    First, applying ray-marching (as in NeRF [38]) with the attenuation-based model should produce the same rendered image as Gaussian Splatting

    work page

  62. [62]

    Second, the transmittance function should reach its maximum value at the point of maximum contribution along the rayx=o+t G o,ww

    work page

  63. [63]

    Therefore, we assume an isosurface of the transmittance function for a single Gaussian should be a portion of an ellipsoid centered at the Gaussian center

    Third,followingthedefinitionoftransmittancefromObjectsasVolumes[39], we interpret transmittance as visibility probability along a ray. Therefore, we assume an isosurface of the transmittance function for a single Gaussian should be a portion of an ellipsoid centered at the Gaussian center. We note that, transmittance having ellipsoidal isosurfaces is not ...

    work page

  64. [64]

    The rendered normal maps created by rasterizing the 3D Gaussian Splatting representation with normalsni

    work page

  65. [65]

    properly

    The image-derivative of the depth. Doing so promotes the value of our normal field at the 0.5-isosurface of the transmittance to match the true normal of the surface. Proof.First, let us start by showing why supervising the rendered normal maps—produced by splatting Gaussians with their learnable normalsni— also supervises the normal fieldN. For a given r...

    work page