Pith. sign in

REVIEW 4 major objections 5 minor 16 references

Manifold-GS claims that adaptive Gaussian splat scenes can be exported as certified hybrid assets—open surface patches, attached appearance Gaussians, and residual splats—whose editing and collision behavior is governed by a varifold mass l

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 01:00 UTC pith:AGZOXTXT

load-bearing objection Honest, carefully scoped asset-layer paper whose certification hinges on an unvalidated covariance-to-surface link and a thin three-scene benchmark; worth refereeing if the author opens the ledger. the 4 major comments →

arxiv 2608.00214 v1 pith:AGZOXTXT submitted 2026-07-31 cs.CV

Manifold-GS: Certified Hybrid Assets via Varifold-Conservative Gaussian Splatting

Boyang Li This is my paper
classification cs.CV
keywords 3D Gaussian splattingvarifoldcertified surface patcheshybrid asset representationsurface reconstructioncollision detectionidentifiabilityDTU benchmark
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that the value of a 3D Gaussian Splatting scene is not only rendered pixels: the adaptive Gaussian cloud can be turned into a structured, editable asset if one separates geometric mass from appearance opacity and treats surface-like Gaussians as samples of an unoriented varifold. On this view, adaptive refinement becomes a mass-transport problem with explicit conservation rules, and only regions where the local covariance is thin and the tangent field is self-consistent are 'certified' as open surface patches. Everything else stays as residual splats instead of being force-closed into a watertight mesh. The paper reports a frozen three-scene benchmark in which certified patches show zero edit leakage, texture round-trip PSNR above 30 dB, and lower collision floater area than standard mesh extraction, at the cost of lower surface coverage. It also states a boundary: local geometric compatibility does not imply that the true surface is identifiable from sparse RGB images alone.

Core claim

Manifold-GS claims that a trained Gaussian splat can be reinterpreted as a discrete unoriented varifold, with each surface-like primitive contributing a positive geometric mass q_i at a position and a tangent 2-plane. The central design rule is that q_i is not opacity: opacity is a rendering parameter, while q_i is a conserved quadrature mass for area-like measure. Under this rule, clone/split/merge/prune operations are framed as transport that preserves zero-th mass and first spatial moment, with bounded-Lipschitz guarantees, so refinement does not silently create or destroy surface area. The pipeline then exports a hybrid asset: confidence-certified open patches as editable geometry, attac

What carries the argument

The central object is the Gaussian varifold surrogate V_G(ϕ)=Σ q_i ϕ(μ_i, P_i), built from each surface-like Gaussian's center μ_i, the tangent 2-plane P_i determined by the two dominant eigenvectors of its covariance, and a separate positive geometric mass q_i. This object carries the argument because it converts a rendering representation into a geometric measure: the paper's claims about refinement conservation, certified patch extraction, source-preserving editing, and conservative collision candidates all follow from treating q_i as a conserved area-like mass rather than identifying it with opacity. A second piece of machinery is the realizability-vs-identifiability decomposition: compa

Load-bearing premise

The certification pipeline rests on treating covariance thinness as a genuine surface detector: when the two largest eigenvalues dominate, the Gaussian is assumed to carry a real tangent plane, and if trained Gaussians are thin for rendering reasons (haze, anti-aliasing, volumetric softness) rather than because they sit on real geometry, then the certified patches and the edit/collision claims lose their geometric referent.

What would settle it

Run the certification export on a deliberately ambiguous scene—for instance, a flat textureless plane seen only from grazing angles, or a volume of semi-transparent haze that renders like thin Gaussians—and compare the certified patch normals and positions against a laser scan or structured-light depth map. If the pipeline certifies patches whose normals are systematically rotated away from the true surface by more than a few degrees, or if the certified collision candidate contains a phantom surface in free space in that region, then the covariance-to-surface premise fails and the certified-a

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If a Gaussian scene is treated as a varifold with separate mass, adaptive refinement no longer changes the induced geometry silently; clone/split/prune can be audited through mass and moment ledgers.
  • Certified open patches can serve as editable geometry with zero patch-defined edit leakage, because source mappings bind edits to certified patch units rather than to nearby vertices.
  • Collision candidates derived from certified patches produce fewer false collision surfaces than watertight or TSDF-style meshes on the same DTU scenes, at lower coverage.
  • The asset survives common post-processing: after 5k-face simplification, floater percentages and coverage remain nearly unchanged and outputs stay manifold-clean.
  • Local realizability diagnostics improve tangent quality, but without depth or calibrated sparse support they do not identify the ground-truth surface; RGB-only compatibility is a precision regularizer, not a reconstruction method.

Where Pith is reading between the lines

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

  • An unstated consequence is that certification thresholds can be tuned per application: high thresholds for collision and physics proxies, lower thresholds for visualization, rather than a single global surface extraction.
  • The varifold mass ledger could be extended to time-varying scenes or streaming updates, where refinement conservation would bound geometric drift as Gaussians are added or removed incrementally.
  • If the covariance-thinness detector is the weak link, a testable extension is to fuse certified-patch extraction with depth sensing: the paper's identifiability analysis predicts that depth-anchored evidence should raise certified coverage and lower patch Chamfer beyond the RGB-only ceiling.
  • Zero edit leakage is only as strong as the patch approximation: externally selected regions matched to certified patches show IoU 0.27–0.53, so coarser semantic editing still needs a segmentation step before certified binding becomes useful.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes Manifold-GS, a layer that converts trained 3D Gaussian Splatting scenes into hybrid assets: certified open surface patches, attached appearance Gaussians, residual splats, source mappings, and conservative collision candidates. The theoretical component interprets surface-like Gaussians as a discrete unoriented varifold with a separate geometric mass q_i, gives conditional forward-consistency and refinement-conservation bounds (Eqs. 6–9), and argues that local realizability diagnostics do not imply sparse-RGB identifiability. The empirical component reports a frozen three-scene DTU benchmark on editing, texture round-trip, collision precision, simplification robustness, and external-region annotations, alongside explicit negative results showing that the unanchored RGB-only loss does not pass its own 1% improvement threshold. The paper is unusually honest about the conditional and preliminary nature of its claims, framing the contribution as a precision–coverage tradeoff rather than a universal surface-reconstruction win.

Significance. If the central asset claim is validated, the varifold-conservative reinterpretation of Gaussian splats would be a useful conceptual and practical step: it separates geometric mass from opacity, provides a principled way to expose certified geometry rather than hallucinating watertight surfaces in unobserved regions, and ships a reproducible frozen protocol with unit tests and explicit PASS/FAIL checks. The negative finding that local realizability is not sparse-RGB identifiability is also valuable as a boundary statement. However, the paper's empirical support is narrow (three DTU scenes, single-run asset metrics), and the meaning of 'certified' is currently stronger than the evidence. The theoretical results are deliberately conditional and partly definitional, so the significance rests primarily on the asset benchmark, which needs tighter geometric validation before the central claim can be accepted.

major comments (4)
  1. [§7.5, Table 4; §7.8] The term 'certified' is load-bearing for the central claim, but the certification is not validated against per-patch ground-truth geometry. On scan24 the certified collision candidate still has 18.33% floater area at the 1% bbox tolerance, and §7.8 itself states that scan105 floaters can be marked supported. No per-patch GT distance distribution or precision-recall curve for certified versus rejected patches is reported. Without such a curve, an external reader cannot tell whether certification selects real surface patches or merely thin appearance blobs. Please report per-patch Chamfer/accuracy (e.g., fraction within 2mm/5mm of DTU GT), broken down by certified/accepted and rejected patches, over a tolerance sweep, and either tighten the certification criterion or rename the concept from 'certified' to a weaker confidence statement.
  2. [§7.3–§7.5] It is unclear which trained Gaussian model is used for the main asset benchmark. §7.3 shows that the unanchored RGB-only loss fails its own 1% matched-improvement threshold, while only the fixed COLMAP-support anchor passes the frozen replication (with scan105 excluded as post-hoc). §7.5 does not state whether the certified asset bundles are exported from the vanilla, compatibility-loss, or anchored model. If the anchor is required, the asset claims should be scoped to that setting. Please state explicitly which model produces the Gaussians used in Tables 3–4 and, if the anchor is required, discuss how general the asset pipeline is to unanchored 3DGS.
  3. [Tables 2–4; §7.5] The central empirical claims rest on one run per scene, with no multiple seeds, confidence intervals, or run-to-run variance. For a benchmark that is described as 'frozen,' the absence of variability reporting makes it impossible to judge whether the observed differences from 2DGS/SuGaR are meaningful. Please provide at least three runs per scene with different random seeds and report mean±std for floater area, coverage, texture PSNR, and external-motion/leak metrics, under the same matched resource schedule. If this is not feasible for compute reasons, say so explicitly and present the results as a pilot, not a final benchmark.
  4. [Eq. (2)–(3), §3, §7.1] The covariance-to-surface map is the foundation of the whole pipeline, but its fidelity is only sanity-checked on an analytic sphere and indirectly via DTU collision metrics. The paper admits in §7.1 that thinness pressure is 'not evidence of manifold reconstruction.' Given that the downstream edit/collision claims depend on certified patches lying on real surfaces, the paper should include a direct test of the covariance-to-surface map on real scenes: compare certified patch positions/normals to DTU GT within local support, and report the fraction of certified patches whose dominant covariance eigenvectors align with the GT tangent within a specified angle threshold. Without this, the asset claims lack a clear geometric referent.
minor comments (5)
  1. [Abstract, §1, §7.5] The word 'certified' is used in the abstract and throughout before the conditional nature is explained. Recommend consistently writing 'confidence-certified' or 'conservatively certified' and defining the certification criteria (covariance thinness, support, compatibility, Fisher support) in one place in §3.
  2. [Eq. (12), Appendix B] The MMD bound uses a Gaussian kernel and requires equal-mass or normalized measures; this normalization should be stated in the main text where Eq. (12) appears, not only in the appendix.
  3. [Appendix A.2, Eq. (9)] The refinement conservation result is an identity under the implemented recentering/inherited-tangent split. The paper is transparent about this, but it would help to state explicitly in §4 that the proposition is a design constraint, not an empirical property of trained 3DGS, and that the unit tests verify the implementation against this definition.
  4. [Tables 3–4] Define 'ext. move', 'phantom', and 'coverage' directly in the table captions. The text explains them later, but a self-contained caption improves readability.
  5. [§7.8] The Fisher diagnostic is called a 'certificate' in §7.8 but the text immediately qualifies it as a local ranking. The wording is contradictory; use 'support score' or 'local evidence score' to avoid implying a formal geometric certificate.

Circularity Check

3 steps flagged

Zero edit leakage and refinement conservation are by construction; the DTU collision/texture evidence is external but the abstract presents the definitional zero-leak as a benchmark result.

specific steps
  1. self definitional [§6 'Frozen asset protocol' (claim echoed in Abstract)]
    "For editing, certified binding is expected to have zero leak by construction, so the informative check is whether a proximity baseline actually leaks across patch boundaries on the same scene."

    The abstract's 'zero patch-defined certified edit leakage' is not an empirical result: certified binding is defined as source-preserving, so an edit to a certified patch cannot move non-certified source mass by construction. The protocol text itself says the zero is expected 'by construction.' The informative comparison is the proximity baseline, which does leak. Thus one of the headline benchmark claims is an identity enforced by the method's own definition, not a measured outcome.

  2. self definitional [§4, after Proposition 1 and Eq. (9)]
    "The implemented conservative split recenters child offsets and inherits the parent tangent, so mass, barycenter, and inherited-tangent moment are preserved deterministically."

    Proposition 1's conclusion (zero-th mass, first spatial moment, and tangent-projector moment preservation) is exactly the set of equalities in Eq. (8) that the implemented split hard-codes. The unit tests therefore verify the code against its own definition: the 'refinement-conservative' guarantee is an engineering invariant enforced by construction, not evidence about the geometry of trained Gaussian scenes. The paper labels the statement conditional, but the implemented operator is built to satisfy those conditions deterministically.

  3. self definitional [§7.1 'Mechanism: covariance and compatibility']
    "Since the projected field is constructed from the same MLS support, this is a diagnostic sanity check rather than a reconstruction result."

    The analytic-sphere experiment reports large reductions in fundamental-form residuals after offline projection, but the projected field is built from the same MLS support used to compute the diagnostics. The residual decrease is then a consequence of the construction, not an independent confirmation that covariance-derived normals correspond to real geometry. The paper explicitly calls this a 'sanity check,' which is transparent, but it is still an in-sample validation that cannot provide external support for the covariance-to-surface map.

full rationale

The paper has genuine independent content: the DTU collision-vs-GT comparisons (Table 4), texture round-trip PSNR, phantom-collision probes, and simplification robustness are evaluated against external GT and baseline meshes and do not reduce to the method's definitions. There are no load-bearing self-citations; the varifold and differential-geometry references are standard external mathematics. However, three places in the evidence chain are circular by construction. First, the abstract's zero patch-defined certified edit leakage is explicitly 'expected ... by construction' in §6: certified binding is source-preserving, so the zero is an identity; the informative result is the leaking proximity baseline. Second, Proposition 1's conservation statement is identical to the equalities the implemented split hard-codes by recentering child offsets and inheriting the parent tangent, so the unit tests verify the implementation against its own definition. Third, the §7.1 fundamental-form 'reduction' is computed on a field constructed from the same MLS support, so the diagnostic cannot fail on that input; the paper honestly labels this a sanity check. The post-hoc exclusion of scan105 from the two-scene replication is disclosed and is a statistical-validity concern rather than a circularity. Overall, the certified-asset claim is only partially empirically grounded: collision/tradeoff behavior is externally supported, while conservation, zero-leak editing, and the realizability sanity check are invariants of the method's own construction.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 2 invented entities

The central claim rests on a small set of domain assumptions: (a) covariances of trained thin Gaussians faithfully indicate real tangent planes (Eq. 2–3); (b) trained clouds admit partitions with controlled sampling/tangent/mass error (Eq. 6), which the paper concedes is not automatic; (c) the observation-coercivity conditions of Appendix C hold where patches are certified; (d) the DTU STL-to-Gaussian-frame mapping via the preprocessing scale mat is correct. The conservation results are identities by construction, so the contribution is the modeling and protocol, not a free-standing theorem. The free parameters are mostly evaluation thresholds (1% tolerance, 50% overlap, 1% rule, 5k faces); the two kernel bandwidths (σ, τ) affect all reported varifold-distance numbers.

free parameters (7)
  • Kernel bandwidths σ, τ = not disclosed
    Eq. 22 varifold MMD; all reported "Varif." distances (Table 1) depend on these hand-chosen scales.
  • Collision tolerance = 1% of GT bbox diagonal
    §7.6 headline floater numbers are quoted at this single tolerance; protocol partially mitigates via a coverage sweep.
  • External-region overlap threshold = ≥50% patch overlap
    §7.5 converts manual Blender selections into certified-patch approximations; IoU/precision/recall numbers depend on it.
  • Matched-improvement threshold = 1% relative overall DTU improvement
    §7.3 pass/fail rule for the three-scene replication; 0.271% fails, two-scene 2.97% passes only after excluding scan105.
  • Simplification face target = ~5k faces
    §7.7 robustness numbers are conditional on this target.
  • Thinness criterion = not specified (λi1, λi2 ≫ λi3)
    §3 gate for "surface-like" status; no numeric cutoff is disclosed in the draft.
  • Fisher support thresholds = unspecified; 19/17/24 weak patches
    §7.8 acceptance counts imply hidden thresholds that determine the "locally supported" certificate.
axioms (6)
  • domain assumption Covariance thinness implies a faithful local tangent plane and surface normal (Σ_i eigensystem, Eq. 2–3).
    The entire certification pipeline (patch acceptance, normals, collision candidates) inherits this map; the only external check is DTU GT collision on three scenes.
  • domain assumption Trained Gaussian clouds admit a partition with controlled sampling, tangent, and mass error (Eq. 6).
    Needed for the Eq. 7 forward-consistency bound; the paper concedes trained 3DGS does not automatically satisfy it (§3).
  • standard math Varifold / geometric measure theory background (Simon [11]).
    Unoriented varifold surrogate V_G is defined via Grassmannian test functions (Definition 1).
  • standard math Fundamental-form compatibility (Gauss/Codazzi) is the correct local realizability test (do Carmo [4]).
    Used in §5 diagnostics: normal curl, Gauss residual, Codazzi-like residual.
  • domain assumption Observation coercivity (Appendix C): depth H1 error controls position+projector; RGB identifiability requires strictly positive smallest singular value of the restricted rendering Jacobian.
    Asserted as the boundary between realizability and identifiability; measured only via indirect Fisher scores (§7.8), not directly on real scenes.
  • domain assumption DTU STL is deterministically mapped to the Gaussian/reconstruction frame by the preprocessing scale mat with no ICP (§7.6).
    All collision/phantom baselines are evaluated in this alignment; a wrong GT mapping invalidates every floater comparison.
invented entities (2)
  • Geometric quadrature mass qi no independent evidence
    purpose: Conserved area-like mass decoupled from opacity; the ledger quantity transported by split/merge/prune.
    Conservation is enforced by construction (§4), and qi is never measured against external area; only aggregate GT collision statistics give indirect support.
  • Certified open surface patch independent evidence
    purpose: Trusted unit of hybrid asset export (editable, collision-relevant), with source mappings and a conservative collision candidate.
    Acceptance uses realizability diagnostics not fitted to DTU floater numbers, so low floater/phantom rates vs external GT are falsifiable predictions; however the certificate itself is a latent construct validated only via three DTU scenes.

pith-pipeline@v1.3.0-alltime-deepseek · 10693 in / 22063 out tokens · 212399 ms · 2026-08-04T01:00:21.336900+00:00 · methodology

0 comments
read the original abstract

3D Gaussian Splatting (3DGS) gives high-quality novel-view synthesis, but its adaptive radiance primitives are not directly usable as structured assets: opacity is not an additive area measure, refinement can change the induced geometry, and watertight mesh extraction can hallucinate collision surfaces in unobserved regions. We introduce Manifold-GS, a certified hybrid asset layer for Gaussian scenes. The method separates appearance opacity from geometric quadrature mass, represents surface-like Gaussians as a discrete unoriented varifold, and exports only confidence-certified open surface patches while retaining uncertified content as residual splats. It provides refinement-conservative mass transport, local realizability diagnostics, source-preserving patch bindings, and conservative collision candidates. On three DTU scenes, a frozen asset benchmark shows zero patch-defined edit leakage, texture round-trip PSNR of 30.1/35.3/33.7 dB, and lower collision floater area than official 2DGS meshes on all scenes, with large gaps on two scenes. The result is a precision-coverage tradeoff rather than a universal reconstruction claim. External-region annotations, phantom-collision probes, and 5k-face simplification further support the certified asset interpretation, while RGB-only experiments show that local realizability is not sufficient for sparse-view surface identifiability.

Figures

Figures reproduced from arXiv: 2608.00214 by Boyang Li.

Figure 1
Figure 1. Figure 1: Collision cross-section demonstration on DTU scan24. The row visualizes the GT surface together with candidate [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Certified edit propagation on scan105. A selected part is transformed through source-preserving certified patch binding, [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

16 extracted references · 5 linked inside Pith

  1. [1]

    Meshsplat: Generalizable sparse-view surface reconstruction via gaussian splatting

    Hanzhi Chang, Ruijie Zhu, Wenjie Chang, Mulin Yu, Yanzhe Liang, Jiahao Lu, Zhuoyuan Li, and Tianzhu Zhang. Meshsplat: Generalizable sparse-view surface reconstruction via gaussian splatting. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 2716–2724, 2026

    2026

  2. [2]

    PGSR: Planar-based gaussian splat- ting for efficient and high-fidelity surface reconstruction

    Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. PGSR: Planar-based gaussian splat- ting for efficient and high-fidelity surface reconstruction. arXiv preprint arXiv:2406.06521, 2024

    Pith/arXiv arXiv 2024

  3. [3]

    VCR-GauS: View consis- tent depth-normal regularizer for gaussian surface recon- struction

    Hanlin Chen, Fangyin Wei, Chen Li, Tianxin Huang, Yun- song Wang, and Gim Hee Lee. VCR-GauS: View consis- tent depth-normal regularizer for gaussian surface recon- struction. InAdvances in Neural Information Processing Systems, 2024

    2024

  4. [4]

    do Carmo.Differential Geometry of Curves and Surfaces

    Manfredo P. do Carmo.Differential Geometry of Curves and Surfaces. Prentice-Hall, 1976

    1976

  5. [5]

    SuGaR: Surface- aligned gaussian splatting for efficient 3d mesh recon- struction and high-quality mesh rendering

    Antoine Gu´edon and Vincent Lepetit. SuGaR: Surface- aligned gaussian splatting for efficient 3d mesh recon- struction and high-quality mesh rendering. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024

    2024

  6. [6]

    2d gaussian splatting for geomet- rically accurate radiance fields

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geomet- rically accurate radiance fields. InACM SIGGRAPH Conference Papers, 2024

    2024

  7. [7]

    Multi-view normal and distance guidance gaussian splatting for surface recon- struction.arXiv preprint arXiv:2508.07701, 2025

    Bo Jia, Yanan Guo, Ying Chang, Benkui Zhang, Ying Xie, Kangning Du, and Lin Cao. Multi-view normal and distance guidance gaussian splatting for surface recon- struction.arXiv preprint arXiv:2508.07701, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42(4), 2023

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42(4), 2023

    2023

  9. [9]

    G. Lee, S. Hong, and J. Suh. Prior-driven enhancements in 3d gaussian splatting: Normals and depths regulariza- tion.International Archives of the Photogrammetry, Re- mote Sensing and Spatial Information Sciences, XLVIII- G-2025:891–897, 2025

    2025

  10. [10]

    Sorted opacity fields for fast unbounded sur- face reconstruction

    Lukas Radl, Felix Windisch, Thomas Deixelberger, Jozef Hladky, Michael Steiner, Dieter Schmalstieg, and Markus Steinberger. Sorted opacity fields for fast unbounded sur- face reconstruction. InACM SIGGRAPH Asia Conference Papers, 2025

    2025

  11. [11]

    Australian National University, 1983

    Leon Simon.Lectures on Geometric Measure Theory. Australian National University, 1983

    1983

  12. [12]

    Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians.arXiv preprint arXiv:2403.17898, 2024

    Lu Wang, Juncheng Zhang, Xiyao Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Lan Yu, and Lan Xu. Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians.arXiv preprint arXiv:2403.17898, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    GSurf: Learning signed distance fields from splatting opaque gaussians for high-quality 3d reconstruction.Computer- Aided Design, page 104106, 2026

    Baixin Xu, Jiangbei Hu, Jiaze Li, and Ying He. GSurf: Learning signed distance fields from splatting opaque gaussians for high-quality 3d reconstruction.Computer- Aided Design, page 104106, 2026

    2026

  14. [14]

    Geosplatting: Towards geometry guided gaussian splatting for physically-based inverse rendering

    Kai Ye, Chong Gao, Guanbin Li, Wenzheng Chen, and Baoquan Chen. Geosplatting: Towards geometry guided gaussian splatting for physically-based inverse rendering. InProceedings of the IEEE/CVF International Confer- ence on Computer Vision, pages 28991–29000, 2025

    2025

  15. [15]

    Mip-splatting: Alias-free 3d gaus- sian splatting.arXiv preprint arXiv:2311.16493, 2024

    Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaus- sian splatting.arXiv preprint arXiv:2311.16493, 2024

    Pith/arXiv arXiv 2024

  16. [16]

    Gaus- sian opacity fields: Efficient and compact surface re- construction in unbounded scenes.arXiv preprint arXiv:2404.10772, 2024

    Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaus- sian opacity fields: Efficient and compact surface re- construction in unbounded scenes.arXiv preprint arXiv:2404.10772, 2024. 9

    Pith/arXiv arXiv 2024