pith. sign in

arxiv: 2607.01556 · v1 · pith:UD7FKVHDnew · submitted 2026-07-02 · 💻 cs.CV

Mind the Gap: Standard 3DGS Evaluation Primarily Measures Near-Trajectory Interpolation

Gaoxiang Jia , Vikram Appia This is my paper

Pith reviewed 2026-07-03 00:48 UTC · model grok-4.3

classification 💻 cs.CV
keywords 3D Gaussian Splattingevaluation protocolinterpolation vs extrapolationspatial generalizationview synthesisNeRFholdout setscamera trajectory
0
0 comments X

The pith

Standard 3D Gaussian Splatting evaluation measures near-trajectory interpolation rather than spatial generalization.

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

The paper demonstrates that the common practice of holding out every N-th frame in 3D scene reconstruction benchmarks primarily assesses how well models interpolate between nearby trained views instead of generalizing to distant spatial locations. Using a new protocol that matches the number of training images but varies only whether holdouts are evenly distributed or form a contiguous unseen sector, it reveals a consistent 3 to 12 dB performance drop for the extrapolation case. This gap is larger than differences between competing methods, can change method rankings, and holds across Gaussian and non-Gaussian representations, showing the issue stems from spatial coverage rather than the specific model. Standard holdout metrics therefore indicate near-trajectory rendering quality but do not support claims of broad spatial generalization.

Core claim

Standard MipNeRF360-style 3D Gaussian Splatting evaluation holds out every N-th frame, but these frames have trained neighbors on both sides, so the metric measures near-trajectory interpolation rather than spatial generalization. A matched-count protocol isolating this effect by training on the same number of images but differing in holdout distribution finds a large consistent interpolation-extrapolation gap of 3~12dB that exceeds typical method differences, is robust to training noise, can flip rankings, and persists across representation families including NeRF.

What carries the argument

The matched-count paired holdout protocol, where one arm uses evenly spread holdouts for interpolation and the other uses a contiguous spatial sector for extrapolation, while keeping training image count identical.

If this is right

  • The performance gap is dominated by a diffuse/geometry-proxy component.
  • The gap tracks each view's angular distance to its nearest training view, serving as a zero-cost signal for capture planning.
  • Loss-side regularization produces only marginal gains.
  • Standard holdouts remain useful for near-trajectory rendering but should not alone be interpreted as evidence of spatial generalization.

Where Pith is reading between the lines

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

  • Future benchmarks could routinely report both interpolation and extrapolation metrics to distinguish these capabilities.
  • The angular distance to nearest training view could be incorporated into training objectives to encourage better spatial coverage.
  • This evaluation gap may affect conclusions in other novel view synthesis tasks where camera trajectories are limited.
  • Capture strategies that maximize minimum angular separation might reduce the observed extrapolation penalty.

Load-bearing premise

The contiguous spatial sector holdout and the evenly spread holdout differ only in interpolation versus extrapolation properties and introduce no other systematic differences in scene coverage, lighting, or reconstruction difficulty when the image count is matched.

What would settle it

Finding that performance on the contiguous sector holdout matches the even-spread holdout performance when the number of training images is identical, or that controlling for angular distance eliminates the gap.

Figures

Figures reproduced from arXiv: 2607.01556 by Gaoxiang Jia, Vikram Appia.

Figure 1
Figure 1. Figure 1: Interpolation vs. extrapolation evaluation. (a) Standard evaluation: held-out frames (red) have trained neighbors on both sides—the metric measures interpola￾tion. (b) Spatial-holdout evaluation: a contiguous region is excluded from training. Held-out frames now have no nearby training views—the metric measures extrapolation. Our fair matched-count protocol ensures both arms train on the same number of ima… view at source ↗
Figure 3
Figure 3. Figure 3: Nearest training-view distance and the gap. (a) By construction, extrapolation test views are 4.1× far￾ther from their nearest training view than interpolation views (mean 15.2 ◦ vs. 3.7 ◦ across the 9 scenes). (b) The mechanistic test: per-image PSNR falls with nearest-view distance (n = 492, r = −0.58, p < 10−44), and the trend holds within every scene. Lines connect the two arms of each scene [PITH_FUL… view at source ↗
Figure 4
Figure 4. Figure 4: Gap attribution by SH degree. (a) Interp PSNR improves with SH degree (view-dependent color helps); extrap PSNR gains less. (b) Per-scene decomposition: blue = diffuse/geometry-proxy gap (SH=0), orange = view￾dependent gap. Stars (⋆) mark scenes where SH hurts ex￾trapolation. Mean: 62% diffuse, 38% VD [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
read the original abstract

Standard MipNeRF360-style 3D Gaussian Splatting (3DGS) evaluation holds out every N-th frame -- but these frames have trained neighbors on both sides, so the metric measures near-trajectory interpolation rather than spatial generalization. We introduce a fair matched-count protocol that isolates this effect: both arms train on the same number of images and differ only in whether the holdout is spread evenly (interpolation) or forms a contiguous spatial sector (extrapolation). Our primary finding is a large, consistent interpolation-extrapolation gap of 3~12dB -- several times the differences typically reported between competing methods. The gap is robust to training noise, is in two cases large enough to flip a method ranking under multi-seed confirmation, and -- crucially -- persists across three representation families, including a non-Gaussian volumetric neural radiance field (NeRF), so it reflects spatial coverage rather than any one representation. Diagnostically, it is dominated by a diffuse/geometry-proxy component and tracks each view's angular distance to its nearest training view, a zero-cost signal that also guides capture planning; loss-side regularization yields only marginal gains. Standard holdouts remain useful for near-trajectory rendering but should not, alone, be read as evidence of spatial generalization. Prior work notes protocol sensitivity; ours is, to our knowledge, the first to combine matched-count paired holdout, cross-representation quantification, and a diagnostic analysis Table 1. We describe a spatial-holdout benchmark toolkit with standardized splits and baselines for 16 scenes, which we are preparing for public release.

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 claims that standard MipNeRF360-style 3DGS evaluation (every-Nth-frame holdouts) primarily measures near-trajectory interpolation rather than spatial generalization. It introduces a matched-count protocol that compares evenly-spread holdouts (interpolation) against contiguous spatial-sector holdouts (extrapolation) while fixing training image count, reports a consistent 3-12 dB PSNR gap across three representation families (including NeRF), shows the gap can flip method rankings under multi-seed evaluation, and provides an angular-distance diagnostic that tracks per-view difficulty.

Significance. If the reported gap is attributable solely to interpolation vs. extrapolation properties rather than training-set confounds, the result would be significant for novel-view synthesis evaluation: the gap magnitude exceeds typical inter-method differences, the cross-representation persistence (including non-Gaussian NeRF) strengthens the spatial-coverage interpretation, and the zero-cost angular-distance diagnostic offers a practical capture-planning signal. The multi-seed confirmation and preparation of a 16-scene benchmark toolkit are additional strengths.

major comments (3)
  1. [matched-count protocol (abstract and §3)] The matched-count protocol description does not demonstrate that angular statistics or spatial density of the training views are equalized between the evenly-spread and contiguous-sector arms; removing a contiguous block necessarily creates larger gaps in the training trajectory itself, which could alter optimization dynamics, regularization, or learned proxies independently of the test-set interpolation/extrapolation distinction.
  2. [Table 1 and multi-seed experiments] Table 1 and the multi-seed results report consistent gaps and occasional ranking flips but provide no error bars, standard deviations, or statistical significance tests, weakening the claim that the gap is robust to training noise and large enough to reliably flip rankings.
  3. [diagnostic analysis and §4] While the angular-distance diagnostic is shown to track per-view error, the manuscript does not include a control experiment that holds training-view angular coverage fixed while varying only test-set properties, leaving the central isolation of interpolation vs. extrapolation unverified.
minor comments (2)
  1. [abstract] The abstract states the toolkit is 'preparing for public release' but does not specify the exact split definitions or baseline implementations that will be released, which would aid reproducibility.
  2. [throughout] Notation for the two holdout arms (e.g., 'evenly spread' vs. 'contiguous sector') should be defined once with a short acronym or label for consistent use in figures and tables.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive comments and for recognizing the potential impact of our work. We provide point-by-point responses to the major comments below. We will make revisions to address the concerns raised.

read point-by-point responses

  1. Referee: [matched-count protocol (abstract and §3)] The matched-count protocol description does not demonstrate that angular statistics or spatial density of the training views are equalized between the evenly-spread and contiguous-sector arms; removing a contiguous block necessarily creates larger gaps in the training trajectory itself, which could alter optimization dynamics, regularization, or learned proxies independently of the test-set interpolation/extrapolation distinction.

    Authors: We agree that the manuscript does not explicitly demonstrate equalization of angular statistics or spatial density for the training views. The protocol is designed to fix the training image count while varying the holdout configuration to isolate test-set properties. However, as noted, the contiguous holdout introduces a larger gap in the training trajectory. In the revision, we will add quantitative comparisons of the angular coverage and gap statistics for the training sets in both protocols (e.g., histograms of angular distances between training views). This will clarify the extent to which training dynamics may differ and allow readers to assess the contribution of test-set extrapolation. We will also discuss this as a potential confounding factor. revision: yes

  2. Referee: [Table 1 and multi-seed experiments] Table 1 and the multi-seed results report consistent gaps and occasional ranking flips but provide no error bars, standard deviations, or statistical significance tests, weakening the claim that the gap is robust to training noise and large enough to reliably flip rankings.

    Authors: We acknowledge the absence of error bars and statistical tests in the current manuscript. To strengthen these claims, we will expand the multi-seed experiments, report standard deviations across seeds, and include statistical significance tests (e.g., paired t-tests) for the observed gaps and ranking flips in the revised Table 1 and associated text. This will provide a more rigorous quantification of robustness to training noise. revision: yes

  3. Referee: [diagnostic analysis and §4] While the angular-distance diagnostic is shown to track per-view error, the manuscript does not include a control experiment that holds training-view angular coverage fixed while varying only test-set properties, leaving the central isolation of interpolation vs. extrapolation unverified.

    Authors: The matched-count protocol aims to isolate the effect by fixing training count, but we recognize that a stricter control holding training angular coverage exactly fixed while varying only the test set would provide stronger isolation. Constructing such a control is non-trivial because the holdout selection directly affects both. In the revision, we will add a discussion of this limitation and explore whether additional experiments (e.g., using synthetic view selection to match angular stats) can be included. The persistence of the gap across representations and the diagnostic's correlation with angular distance provide supporting evidence, but we agree a dedicated control would be valuable. revision: partial

Circularity Check

0 steps flagged

No circularity: purely empirical measurement of holdout gap

full rationale

The paper reports an empirical comparison of two matched-count holdout protocols (evenly-spread vs. contiguous-sector) across multiple scenes and representations, measuring PSNR/SSIM differences directly from rendered outputs. No equations, derivations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided text or abstract. The central claim is the observed 3-12 dB gap and its correlation with angular distance; these are data-driven observations, not quantities constructed by definition from the inputs or prior self-work. The protocol description and diagnostic analysis stand independently without reducing to any of the enumerated circular patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The claim that the gap reflects spatial coverage rather than representation-specific artifacts rests on the domain assumption that the two holdout designs isolate only the interpolation/extrapolation distinction when image counts are matched.

axioms (1)
  • domain assumption The contiguous spatial sector holdout and the evenly spread holdout differ solely in spatial distribution and introduce no other systematic differences in reconstruction difficulty when the number of training images is identical.
    This premise is required for the matched-count protocol to isolate the interpolation-extrapolation effect as described in the abstract.

pith-pipeline@v0.9.1-grok · 5823 in / 1440 out tokens · 36433 ms · 2026-07-03T00:48:48.485834+00:00 · methodology

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

57 extracted references · 57 canonical work pages · 3 internal anchors

  1. [1]

    Kerbl, G

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis. 3D Gaussian Splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4), 2023

    work page 2023

  2. [2]

    HY-World 2.0: A Multi-Modal World Model for Reconstructing, Generating, and Simulating 3D Worlds

    HunyuanWorld Team. HY-World 2.0: A multi-modal world model for reconstructing, generating, and simulating 3D worlds.arXiv preprint arXiv:2604.14268, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  3. [3]

    H.-X. Yu, H. Duan, C. Herrmann, W. T. Freeman, and J. Wu. WonderWorld: Interactive 3D scene generation from a single image. InCVPR, 2025

    work page 2025

  4. [4]

    Gao,et al.CAT3D: Create anything in 3D with multi- view diffusion models

    R. Gao,et al.CAT3D: Create anything in 3D with multi- view diffusion models. InNeurIPS, 2024

    work page 2024

  5. [5]

    Yu,et al.ViewCrafter: Taming video diffusion models for high-fidelity novel view synthesis.IEEE TPAMI, 2025

    W. Yu,et al.ViewCrafter: Taming video diffusion models for high-fidelity novel view synthesis.IEEE TPAMI, 2025

    work page 2025

  6. [6]

    J. T. Barron, B. Mildenhall, D. Verbin, P. P. Srinivasan, and P. Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. InCVPR, 2022

    work page 2022

  7. [7]

    Knapitsch, J

    A. Knapitsch, J. Park, Q.-Y . Zhou, and V . Koltun. Tanks and Temples: Benchmarking large-scale scene reconstruc- tion.ACM Trans. Graph., 36(4), 2017

    work page 2017

  8. [8]

    Martin-Brualla, N

    R. Martin-Brualla, N. Radwan, M. S. M. Sajjadi,et al.NeRF in the Wild: Neural radiance fields for unconstrained photo collections. InCVPR, 2021

    work page 2021

  9. [9]

    V . Ye, R. Li, J. Kerr, M. Turkulainen,et al.gsplat: An open- source library for Gaussian Splatting.JMLR, 26(34), 2025

    work page 2025

  10. [10]

    Kulhanek and T

    J. Kulhanek and T. Sattler. NerfBaselines: Consistent and reproducible evaluation of novel view synthesis methods. In NeurIPS Datasets & Benchmarks, 2025

    work page 2025

  11. [11]

    Ye,et al.AbsGS: Recovering fine details for 3D Gaussian Splatting

    Z. Ye,et al.AbsGS: Recovering fine details for 3D Gaussian Splatting. InACM MM, 2024

    work page 2024

  12. [12]

    Yu,et al.Mip-Splatting: Alias-free 3D Gaussian Splat- ting

    Z. Yu,et al.Mip-Splatting: Alias-free 3D Gaussian Splat- ting. InCVPR, 2024

    work page 2024

  13. [13]

    J. C. Lee, D. Rho, X. Sun, J. H. Ko, and E. Park. Compact 3D Gaussian representation for radiance field. InCVPR, 2024

    work page 2024

  14. [14]

    Huang, Z

    B. Huang, Z. Yu, A. Chen, A. Geiger, and S. Gao. 2D Gaus- sian Splatting for geometrically accurate radiance fields. In SIGGRAPH, 2024

    work page 2024

  15. [15]

    Liang, X

    Y . Liang, X. Yang, J. Lin, H. Li, X. Xu, and Y . Chen. Lu- cidDreamer: Towards high-fidelity text-to-3D generation via interval score matching. InCVPR, 2024

    work page 2024

  16. [16]

    Sargent, Z

    K. Sargent, Z. Li, T. Shah, C. Herrmann, H.-X. Yu, Y . Zhang, et al.ZeroNVS: Zero-shot 360-degree view synthesis from a single image. InCVPR, 2024

    work page 2024

  17. [17]

    Mildenhall, P

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. InECCV, 2020

    work page 2020

  18. [18]

    Warburg, E

    F. Warburg, E. Weber, M. Tancik, A. Holynski, and A. Kanazawa. Nerfbusters: Removing ghostly artifacts from casually captured NeRFs. InICCV, 2023

    work page 2023

  19. [19]

    X. Chen, Q. Zhang, X. Li, Y . Chen, Y . Feng, X. Wang, and J. Wang. Hallucinated neural radiance fields in the wild. In CVPR, 2022

    work page 2022

  20. [20]

    Kotovenko, O

    D. Kotovenko, O. Grebenkova, and B. Ommer. EDGS: Elim- inating densification for efficient convergence of 3DGS. In CVPR, 2026

    work page 2026

  21. [21]

    Han,et al.Extrapolated urban view synthesis benchmark

    X. Han,et al.Extrapolated urban view synthesis benchmark. InICCV, 2025

    work page 2025

  22. [22]

    Hwang, M.-J

    S. Hwang, M.-J. Kim, T. Kang, J. Kang, and J. Choo. VEGS: View Extrapolation of Urban Scenes in 3D Gaussian Splat- ting using Learned Priors. InECCV, 2024

    work page 2024

  23. [23]

    SLAM&Render: A Benchmark for the Intersection Between Neural Rendering, Gaussian Splatting and SLAM

    S. Cerezo, G. Meli, T. Berriel Martins, K. Safronov, and J. Civera. SLAM&Render: A benchmark for the inter- section of neural rendering and SLAM.arXiv preprint arXiv:2504.13713, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  24. [24]

    Xiong,et al.ICO-GS: Intrinsic geometry-appearance consistency optimization for sparse-view Gaussian Splatting

    K. Xiong,et al.ICO-GS: Intrinsic geometry-appearance consistency optimization for sparse-view Gaussian Splatting. arXiv preprint arXiv:2603.02893, 2026

    work page arXiv 2026

  25. [25]

    Xu,et al.DropoutGS: Dropping out Gaussians for better sparse-view rendering

    X. Xu,et al.DropoutGS: Dropping out Gaussians for better sparse-view rendering. InCVPR, 2025

    work page 2025

  26. [26]

    Xing,et al.3DGS-IEval-15K: A large-scale image quality evaluation database for 3D Gaussian Splatting

    Y . Xing,et al.3DGS-IEval-15K: A large-scale image quality evaluation database for 3D Gaussian Splatting. InACM MM, 2025

    work page 2025

  27. [27]

    Geo-EVS: Geometry-Conditioned Extrapolative View Synthesis for Autonomous Driving

    Y . Lan,et al.Geo-EVS: Geometry-conditioned extrapola- tive view synthesis for autonomous driving.arXiv preprint arXiv:2604.07250, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  28. [28]

    Tan,et al.ExtraGS: Geometric-aware trajectory ex- trapolation with uncertainty-guided generative priors.arXiv preprint arXiv:2508.15529, 2025

    K. Tan,et al.ExtraGS: Geometric-aware trajectory ex- trapolation with uncertainty-guided generative priors.arXiv preprint arXiv:2508.15529, 2025. 9

    work page arXiv 2025

  29. [29]

    Paliwal,et al.CoherentGS: Sparse novel view synthesis with coherent 3D Gaussians

    A. Paliwal,et al.CoherentGS: Sparse novel view synthesis with coherent 3D Gaussians. InECCV, 2024

    work page 2024

  30. [30]

    Kheradmand, D

    S. Kheradmand, D. Rebain, G. S. Sharma, W. Sun, Y . C. Tseng, H. Isack, A. Kar, A. Tagliasacchi, and K. M. Yi. 3D Gaussian Splatting as Markov chain Monte Carlo. In NeurIPS, 2024

    work page 2024

  31. [31]

    T. Lu, M. Yu, L. Xu, Y . Xiangli, L. Wang, D. Lin, and B. Dai. Scaffold-GS: Structured 3D Gaussians for view- adaptive rendering. InCVPR, 2024

    work page 2024

  32. [32]

    Fang,et al.Mini-Splatting: Representing scenes with a constrained number of Gaussians

    G. Fang,et al.Mini-Splatting: Representing scenes with a constrained number of Gaussians. InECCV, 2024

    work page 2024

  33. [33]

    Cheng,et al.GaussianPro: 3D Gaussian Splatting with progressive propagation

    K. Cheng,et al.GaussianPro: 3D Gaussian Splatting with progressive propagation. InICML, 2024

    work page 2024

  34. [34]

    Wang,et al.FreeSplat: Generalizable 3D Gaussian Splatting towards free-view synthesis of indoor scenes

    Y . Wang,et al.FreeSplat: Generalizable 3D Gaussian Splatting towards free-view synthesis of indoor scenes. In NeurIPS, 2024

    work page 2024

  35. [35]

    Charatan, S

    D. Charatan, S. Li, A. Tagliasacchi, and V . Sitzmann. pix- elSplat: 3D Gaussian Splats from image pairs for scalable generalizable 3D reconstruction. InCVPR, 2024

    work page 2024

  36. [36]

    Y . Chen, H. Xu, C. Zheng, B. Zhuang, M. Pollefeys, A. Geiger, T. H. Cai, and J. Liang. MVSplat: Efficient 3D Gaussian Splatting from sparse multi-view images. In ECCV, 2024

    work page 2024

  37. [37]

    H. Xu, B. Peng, Z. Cai, H. Zhou, B. Zhuang, and S. Soatto. DepthSplat: Connecting Gaussian Splatting and Depth. In CVPR, 2025

    work page 2025

  38. [38]

    arXiv preprint arXiv:2403.20309 (2024) 2, 3

    Z. Fan,et al.InstantSplat: Unbounded sparse-view pose- free Gaussian Splatting in 40 seconds.arXiv preprint arXiv:2403.20309, 2024

    work page arXiv 2024

  39. [39]

    Turkulainen,et al.DN-Splatter: Depth and normal priors for Gaussian Splatting and meshing

    M. Turkulainen,et al.DN-Splatter: Depth and normal priors for Gaussian Splatting and meshing. InWACV, 2025

    work page 2025

  40. [40]

    Hanson, A

    A. Hanson, A. Tu, V . Singla, M. Jayawardhana, M. Zwicker, and T. Goldstein. PUP 3D-GS: Principled uncertainty prun- ing for 3D Gaussian Splatting. InCVPR, 2025

    work page 2025

  41. [41]

    Kulhanek, S

    J. Kulhanek, S. Peng, Z. Kukelova, M. Pollefeys, and T. Sat- tler. WildGaussians: 3D Gaussian Splatting in the wild. In NeurIPS, 2024

    work page 2024

  42. [42]

    Q. Wang, L. Fan, Y . Wang, Y . Chen, and Z. Zhang. FreeVS: Generative view synthesis on free driving trajectory. In ICLR, 2025

    work page 2025

  43. [43]

    W. Xiao, R. Santa Cruz, D. Ahmedt-Aristizabal, O. Salvado, C. Fookes, and L. Lebrat. NeRF Director: Revisiting view selection in neural volume rendering. InCVPR, 2024

    work page 2024

  44. [44]

    arXiv preprint arXiv:2411.14208 (2024)

    K. Liu,et al.ViewExtrapolator: Novel view ex- trapolation with video diffusion priors.arXiv preprint arXiv:2411.14208, 2024

    work page arXiv 2024

  45. [45]

    M. Shih, W. Ma,et al.ExtraNeRF: Visibility-aware view extrapolation of neural radiance fields with diffusion models. InCVPR, 2024

    work page 2024

  46. [46]

    Niemeyer, J

    M. Niemeyer, J. T. Barron, B. Mildenhall, M. S. M. Sajjadi, A. Geiger, and N. Radwan. RegNeRF: Regularizing neu- ral radiance fields for view synthesis from sparse inputs. In CVPR, 2022

    work page 2022

  47. [47]

    H. Park, G. Ryu, and W. Kim. DropGaussian: Structural regularization for sparse-view Gaussian Splatting. InCVPR, 2025

    work page 2025

  48. [48]

    Yun,et al.FASR: Frequency-adaptive sharpness regular- ization for improving 3D Gaussian Splatting generalization

    Y . Yun,et al.FASR: Frequency-adaptive sharpness regular- ization for improving 3D Gaussian Splatting generalization. arXiv preprint arXiv:2511.17918, 2025

    work page arXiv 2025

  49. [49]

    K. Chen, Y . Zhong, Z. Li, J. Lin, Y . Chen, M. Qin, and H. Wang. Quantifying and alleviating co-adaptation in sparse-view 3D Gaussian Splatting. InNeurIPS, 2025

    work page 2025

  50. [50]

    J. Ke, Q. Wang, Y . Wang, P. Milanfar, and F. Yang. MUSIQ: Multi-scale image quality transformer. InICCV, 2021

    work page 2021

  51. [51]

    L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, and H. Zhao. Depth Anything V2. InNeurIPS, 2024

    work page 2024

  52. [52]

    Fang,et al.Dropping Anchor and Spherical Harmonics for Sparse-view Gaussian Splatting

    G. Fang,et al.Dropping Anchor and Spherical Harmonics for Sparse-view Gaussian Splatting. InCVPR, 2026

    work page 2026

  53. [53]

    Gottwald,et al.PRIMU: Uncertainty estimation for novel views in Gaussian Splatting from primitive-based representations of error and coverage.arXiv preprint arXiv:2508.02443, 2025

    T. Gottwald,et al.PRIMU: Uncertainty estimation for novel views in Gaussian Splatting from primitive-based representations of error and coverage.arXiv preprint arXiv:2508.02443, 2025

    work page arXiv 2025

  54. [54]

    Jiang,et al.FisherRF: Active view selection and uncer- tainty quantification for radiance fields using Fisher infor- mation

    J. Jiang,et al.FisherRF: Active view selection and uncer- tainty quantification for radiance fields using Fisher infor- mation. InECCV, 2024

    work page 2024

  55. [55]

    Chen,et al.COVER: Coverage optimization for camera view selection

    J. Chen,et al.COVER: Coverage optimization for camera view selection. InCVPR, 2026

    work page 2026

  56. [56]

    M ¨uller, A

    T. M ¨uller, A. Evans, C. Schied, and A. Keller. Instant neu- ral graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1–102:15, 2022

    work page 2022

  57. [57]

    R. Li, M. Tancik, and A. Kanazawa. NerfAcc: A general NeRF acceleration toolbox.arXiv preprint arXiv:2210.04847, 2022. 10 Appendix: Supplementary Material This appendix provides multi-metric results, multi-seed val- idation details, method configurations, and an artifact mani- fest. All results use thefair matched-count protocol(both arms train onN−Kimage...

    work page arXiv 2022