Pith. sign in

REVIEW 4 major objections 6 minor 20 references

Matching rendered image edges to SAM3 building edges forces 3D Gaussians into sharp architectural boundaries without changing the 3DGS pipeline.

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 · grok-4.5

2026-07-11 23:21 UTC pith:Z2MTTBY6

load-bearing objection Clean, low-engineering edge loss for 3DGS buildings that modestly improves appearance; the 3D-geometry claim is under-measured. the 4 major comments →

arxiv 2607.03872 v1 pith:Z2MTTBY6 submitted 2026-07-04 cs.CV

SharpSplat: Edge-Regularized 3D Gaussian Splatting for High Fidelity Urban Building Reconstruction from UAV images

Porus Vaid , Shivam Chopra , Vaibhav Kumar This is my paper
classification cs.CV
keywords 3D Gaussian Splattingedge supervisionSAMbuilding reconstructionUAV imagerylarge scene reconstructiondigital twins
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.

Standard 3D Gaussian Splatting blurs building facades, windows, and roof corners because its primitives have smooth falloff and the usual photometric loss never explicitly rewards crisp boundaries. This paper shows that the blur can be corrected by a simple extra loss: extract building-only edges from UAV photos with text-prompted SAM3 masks plus Sobel filtering, then force every rendered training view to match those edges under an L1 gradient alignment. The Gaussians rearrange themselves in 3D so that architectural lines stay sharp from any viewpoint. The method needs no new primitive types, no depth maps, and no changes to the rasterizer, yet it measurably improves edge fidelity on campus, dense-urban, and residential UAV datasets while leaving ordinary image metrics intact or slightly better.

Core claim

Semantic edge regularization—aligning Sobel gradients of rendered views to SAM3-masked building edges via an L1 loss—produces significantly sharper architectural boundaries in 3DGS reconstructions from UAV images without any architectural modification to the 3DGS pipeline.

What carries the argument

The edge alignment loss L_edge: the mean absolute difference between the Sobel edge map of a rendered image and the precomputed SAM3-filtered edge map of the corresponding training photo; added to the ordinary RGB+SSIM objective with a small weight.

Load-bearing premise

That making 2D image-space edge magnitudes match at SAM3-derived building pixels is a reliable enough proxy to force the underlying 3D Gaussians into truly sharp structural geometries, and that the masks themselves are accurate across varied urban styles.

What would settle it

On a held-out UAV building scene, measure whether the reconstructed mesh or novel-view edge maps show measurably sharper roof lines and window frames after the edge loss is added; if edge sharpness stays the same or degrades while photometric scores stay flat, the claim fails.

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

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 / 6 minor

Summary. The paper proposes SharpSplat, a semantic edge-regularization add-on to standard 3D Gaussian Splatting for UAV-based urban building reconstruction. SAM3 text-prompted masks isolate building regions; Sobel gradients on the training images are masked to form E_SAM (Eqs. 1–2). During 3DGS optimization an L1 edge-alignment loss L_edge (Eq. 4) matches Sobel gradients of the rendered image to E_SAM and is added to the usual RGB+SSIM objective (Eq. 5). The authors claim this forces Gaussians into crisp architectural geometries without any change to the 3DGS representation or rasterizer. Quantitative results on three UAV scenes (PolyTech, Art Sci, Gehukheda) at 7k and 15k iterations show small, consistent gains in PSNR/SSIM/LPIPS (Tables I–II); qualitative comparisons (Figs. 2–4) highlight sharper façades and windows versus 3DGS, 2DGS and SuGaR.

Significance. If the edge-fidelity claim is substantiated, the method is a practically useful, architecture-preserving regularizer for digital-twin pipelines that already rely on 3DGS. Strengths include offline SAM3 precomputation, zero extra memory at training time, and evaluation across campus, dense-urban and residential UAV collections. The idea is complementary to surface-oriented methods (2DGS, PG-SAG) that constrain geometry rather than appearance. The technical novelty is incremental—the contribution is essentially a carefully engineered supervisory recipe rather than a new representation—but such lightweight losses can still be valuable if their benefit is cleanly measured.

major comments (4)
  1. Abstract, §IV.B and Tables I–II: the central claim is “significant improvements in edge fidelity” and that L_edge forces “sharp structural geometries.” The only quantitative evidence is global photometric deltas of roughly +0.1–0.2 dB PSNR, +0.004 SSIM and −0.007 LPIPS. These metrics do not isolate edges; without an edge-specific score (gradient correlation, edge F-score vs held-out edges, multi-view façade-landmark error, or similar) the language of “significant edge fidelity” is not supported by the tables.
  2. §III.C, Eq. (4) and the axiom that 2D Sobel L1 matching rearranges 3D means/covariances into sharp multi-view structure: this proxy is asserted but never tested with any 3D geometric measure (surface-normal consistency, mesh-edge sharpness after extraction, or novel-view edge consistency beyond training cameras). Qualitative circles in Figs. 3–4 therefore carry almost the entire weight of the structural-geometry claim and are insufficient for a load-bearing conclusion.
  3. §IV.A and Limitations: λ_edge is acknowledged to require per-scene tuning (range 0.001–0.1) yet no ablation table, sensitivity plot, or default value used for the reported runs is provided. Likewise there is no comparison against a non-semantic baseline (plain image Sobel loss without the SAM3 mask). Without these controls it is impossible to judge how much of the small photometric gain is due to semantic filtering versus generic edge sharpening, or how stable the method is.
  4. Experimental reporting (Tables I–II): results are single-run point estimates with no error bars, multi-seed variance, or statistical test. Given that the absolute gains sit near typical 3DGS run-to-run noise, the consistency claim across datasets cannot be assessed from the published numbers alone.
minor comments (6)
  1. Throughout: “UA V” is consistently written with a space; should be “UAV”.
  2. §IV.A: “Gehukhedaarea” needs a space; resolution factor “1/2” should be stated as half-resolution or 0.5× for clarity.
  3. §III.C / Eq. (5): the numerical value(s) of λ_edge actually used for Tables I–II and Figs. 2–4 are never reported.
  4. Fig. 1 caption and §III.B: rendered edges omit bilateral filtering “by construction,” but the text does not confirm that the same grayscale conversion and kernel normalization are applied to both E_render and E_SAM before the L1.
  5. Related work: several citations are listed as “Anonymous / arXiv preprint, 2025”; if they remain unpublished they should be clearly marked as concurrent or removed from direct quantitative comparison claims.
  6. Limitations paragraph correctly flags SAM3 dependence; a short quantitative note (e.g., mask IoU range on the three datasets, or a failure-case figure) would make the caveat actionable.

Circularity Check

0 steps flagged

No significant circularity: edge loss is external SAM3/Sobel supervision added to 3DGS, not a quantity defined by or fitted to the claimed sharpness result.

full rationale

The paper is an empirical methods contribution that introduces an additive edge-alignment term L_edge = (1/HW) Σ |E_render − E_SAM| (Eq. 4) into the ordinary 3DGS objective L_total = L_RGB + λ_SSIM L_SSIM + λ_edge L_edge (Eq. 5). E_SAM is pre-computed offline from the training images by text-prompted SAM3 masks followed by Sobel filtering (Eqs. 1–2); E_render is obtained by the identical Sobel operator on the differentiable rasterizer output. This supervisory signal is independent of the photometric loss and of the evaluation metrics; no parameter is fitted to the reported PSNR/SSIM/LPIPS numbers and then re-presented as a prediction, nor is any quantity defined in terms of the very 3D edge sharpness it claims to produce. The reference list contains no load-bearing self-citations that supply a uniqueness theorem or an ansatz later treated as external fact. Consequently the derivation chain does not reduce to its own inputs by construction. Any remaining doubts about whether 2D gradient matching is a sufficient proxy for 3D structural geometry are questions of empirical support, not circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The central claim rests on a small set of free hyperparameters (especially λ_edge), standard 3DGS/COLMAP machinery, and the domain assumption that SAM3 text masks plus 2D Sobel edges are a faithful enough proxy for 3D architectural sharpness. No new physical entities are postulated; the main invented construct is the semantic edge map and its L1 alignment loss.

free parameters (4)
  • λ_edge (edge-loss weight) = 0.001–0.1 (per-scene)
    Explicitly stated to require per-scene tuning; authors report values cluster around 0.001–0.1. Directly scales the term that produces the claimed edge improvement.
  • morphological dilation kernel on SAM mask = 5×5
    Fixed 5×5 dilation applied before masking edges; changes which pixels receive edge supervision near boundaries.
  • λ_SSIM = 0.2
    Standard 3DGS photometric weight kept at 0.2; still a free coefficient in the total loss that co-determines the reported metrics.
  • training length (7k / 15k iterations) and resolution factor 1/2 = 7k and 15k iters; res 1/2
    Chosen evaluation budgets that define the reported tables; adaptive density control and learning-rate schedule follow default 3DGS but are not re-ablated with the new loss.
axioms (4)
  • domain assumption SAM3 with the listed text prompts produces binary building masks accurate enough that edges inside those masks are architecturally meaningful for UAV urban scenes.
    Invoked in §III.A and Limitations; entire supervision signal is filtered by M from SAM3.
  • domain assumption Sobel gradient magnitude on RGB (after bilateral filtering on real images) is a suitable differentiable proxy for the sharp boundaries that should appear in the reconstructed model.
    §III.A–B; both E_SAM and E_render are defined via the same Sobel operators.
  • ad hoc to paper Minimizing L1 distance between rendered and SAM-masked edge maps rearranges 3D Gaussians so that structural geometry becomes sharper, not merely so that 2D edges look better from training views.
    Core causal claim in Abstract and §I; only 2D appearance is supervised, yet the paper asserts 3D geometric convergence.
  • domain assumption Standard 3DGS optimization (adaptive densification, exponential LR decay, COLMAP poses) remains valid when the extra edge term is added.
    §III.C and Experimental Setup; no architectural changes claimed.
invented entities (1)
  • Semantic edge map E_SAM and edge alignment loss L_edge no independent evidence
    purpose: Provide the only new supervisory channel that is claimed to force sharp building boundaries.
    Defined in Eqs. (2)–(4); not a physical entity but the paper’s central methodological construct. No independent evidence outside the reported photometric/qualitative results.

pith-pipeline@v1.1.0-grok45 · 10886 in / 3500 out tokens · 35169 ms · 2026-07-11T23:21:05.809742+00:00 · methodology

0 comments
read the original abstract

Reconstructing high-fidelity 3D building models from UAV imagery is essential for large-scale digital twin development. However, existing 3D Gaussian Splatting (3DGS) techniques often struggle with building facades, failing to capture sharp geometric transitions. To address this, we propose a semantic edge regularization framework that supervises 3DGS to produce crisp architectural boundaries. Our method leverages SAM 3 to generate precise building masks, from which we extract architecturally significant edges. During training, we align rendered image gradients with these extracted edges, forcing the Gaussians to converge into sharp structural geometries. Evaluations across campus environments, dense urban centers, and custom residential datasets demonstrate significant improvements in edge fidelity without requiring architectural modifications to the 3DGS pipeline. Our approach proves robust across diverse building types, roof geometries, and urban densities.

Figures

Figures reproduced from arXiv: 2607.03872 by Porus Vaid, Shivam Chopra, Vaibhav Kumar.

Figure 1
Figure 1. Figure 1: Overview of our semantic edge supervision approach. Multi-view images are processed through SAM3 to extract building masks using text prompts. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison with SOTA methods across three scenes. Top row: Art Sci scene. Middle row: PolyTech scene. Bottom row: Gehukheda. From left [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: PolyTech scene comparison. Left: Baseline 3DGS with blurred edges [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Art Sci scene comparison. Left: Baseline 3DGS. Right: Our method. [PITH_FULL_IMAGE:figures/full_fig_p004_4.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

20 extracted references · 7 linked inside Pith

  1. [1]

    Structure-from-motion revisited,

    J. L. Sch ¨onberger and J.-M. Frahm, “Structure-from-motion revisited,” inCVPR, 2016

    2016

  2. [2]

    Nerf: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ra- mamoorthi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,” inECCV, 2020

    2020

  3. [3]

    3d gaussian splatting for real-time radiance field rendering,

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis, “3d gaussian splatting for real-time radiance field rendering,” in ACM SIGGRAPH, 2023

    2023

  4. [4]

    City- gaussianv2: Efficient and geometrically accurate reconstruction for large-scale scenes,

    Y . Xu, J. Guo, Y . Shi, Y . Sun, J. Li, T. Guo, and G. Zhang, “City- gaussianv2: Efficient and geometrically accurate reconstruction for large-scale scenes,”arXiv preprint arXiv:2411.00771, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    Vast- gaussian: Vast 3d gaussians for large scene reconstruction,

    J. Lin, Z. Lin, X. Li, H. Bao, X. Guo, and X. Bao, “Vast- gaussian: Vast 3d gaussians for large scene reconstruction,” in CVPR, 2024

    2024

  6. [6]

    Gs4buildings: Semantic gaussian splatting for building reconstruction,

    W. Zhang, Y . Liu, H. Chen, and Y . Wang, “Gs4buildings: Semantic gaussian splatting for building reconstruction,”ISPRS Journal of Photogrammetry and Remote Sensing, 2025

    2025

  7. [7]

    Det-gs: Depth and edge-aware 3d gaussian splat- ting,

    Anonymous, “Det-gs: Depth and edge-aware 3d gaussian splat- ting,”arXiv preprint, 2025

    2025

  8. [8]

    Dn-splatter: Depth and normal priors for gaussian splatting,

    M. Turkulainen, X. Hua, and C. Ren, “Dn-splatter: Depth and normal priors for gaussian splatting,” inWACV, 2025

    2025

  9. [9]

    Edgegaus- sians: Edge-aware 3d gaussian primitives for scene reconstruc- tion,

    J. Lei, Z. Xu, Q. Hu, H. Wu, Y . Wang, and M. Liu, “Edgegaus- sians: Edge-aware 3d gaussian primitives for scene reconstruc- tion,”arXiv preprint arXiv:2409.09456, 2024

    Pith/arXiv arXiv 2024

  10. [10]

    Sam 3: Segment anything with concepts,

    T. Ju, Y . Wang, W. Liu, Y . Huang, Y . Zhang, and J. Chen, “Sam 3: Segment anything with concepts,”arXiv preprint arXiv:2511.16719, 2025

    Pith/arXiv arXiv 2025

  11. [11]

    Mip- splatting: Alias-free 3d gaussian splatting,

    Z. Yu, A. Chen, B. Huang, T. Sattler, and A. Geiger, “Mip- splatting: Alias-free 3d gaussian splatting,”CVPR, 2024

    2024

  12. [12]

    2d gaussian splatting for geometrically accurate radiance fields,

    B. Huang, Z. Yu, A. Chen, A. Geiger, and S. Gao, “2d gaussian splatting for geometrically accurate radiance fields,” SIGGRAPH, 2024

    2024

  13. [13]

    Citygaussian: Real-time high-quality large-scale scene rendering with gaussians,

    Y . Liu, Y . He, J. Wang, T. Chen, J. Tan, J. Cao, H. Wu, and J. Wang, “Citygaussian: Real-time high-quality large-scale scene rendering with gaussians,”arXiv preprint arXiv:2404.01133, 2024

    Pith/arXiv arXiv 2024

  14. [14]

    Pg-sag: Parallel optimization of 3d gaussians for building reconstruction,

    Anonymous, “Pg-sag: Parallel optimization of 3d gaussians for building reconstruction,”arXiv preprint, 2025

    2025

  15. [15]

    Sf-recon: Large- scale urban scene reconstruction with surface-aware 3d gaussian splatting,

    B. Xiong, Z. Wang, X. Yang, and Z. Li, “Sf-recon: Large- scale urban scene reconstruction with surface-aware 3d gaussian splatting,”arXiv preprint arXiv:2410.05723, 2024

    Pith/arXiv arXiv 2024

  16. [16]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo et al., “Segment anything,” inICCV, 2023

    2023

  17. [17]

    Saga: Segment any 3d gaussians,

    J. Cen, J. Zhou, C. Fang, W. Shen, X. Yang, and Q. Dou, “Saga: Segment any 3d gaussians,”AAAI, 2025

    2025

  18. [18]

    Gaussian grouping: Segment and edit anything in 3d scenes,

    M. Ye, M. Danelljan, F. Yu, and L. Ke, “Gaussian grouping: Segment and edit anything in 3d scenes,” inECCV, 2024

    2024

  19. [19]

    Sa-gs: Semantic-aware gaussian splatting for large scene reconstruction with geometry constrain,

    B. Cen, B. Xiong, Y . Zhang, Z. Zhan, X. Wang, and W. Wu, “Sa-gs: Semantic-aware gaussian splatting for large scene reconstruction with geometry constrain,”arXiv preprint arXiv:2405.16923, 2024

    Pith/arXiv arXiv 2024

  20. [20]

    Urbanscene3d: A large scale urban scene dataset and simulator,

    Y . Lin, J. Wang, D. Held, and D. Ramanan, “Urbanscene3d: A large scale urban scene dataset and simulator,”arXiv preprint arXiv:2107.04286, 2021

    Pith/arXiv arXiv 2021