Pith. sign in

REVIEW 5 major objections 5 minor 45 references

SVR-GS: Spatially Variant Regularization for Probabilistic Masks in 3D Gaussian Splatting

T0 review · 5 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read A spatially variant, per-pixel mask regularizer prunes 3D Gaussian Splatting scenes to about one-fifth of their Gaussians while keeping PSNR loss under half a decibel.

desk verdict Plausible incremental extension of MaskGS with a clean gradient analysis; the compression claim is real but overstated and the importance proxy is unvalidated. read the letter →

arxiv 2509.11116 v1 pith:MGB6UWT7 submitted 2025-09-14 cs.CV

classification cs.CV
keywords 3DGaussiansplattingnovelviewsynthesispruningprobabilisticmaskspatiallyvariantregularizationvisibility-weightedopacityCUDArasterizersparsity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that the standard way of regularizing mask-based pruning in 3D Gaussian Splatting—penalizing the global mean of per-Gaussian existence probabilities—is misaligned with how image quality is actually determined, which is per pixel and per ray. SVR-GS replaces that global penalty with a rendered spatial mask that is large wherever a ray contains many low-importance, occluded, or faint Gaussians. Minimizing the squared energy of this spatial mask pushes the probabilistic masks of those low-importance Gaussians down while sparing the few high-opacity foreground Gaussians that explain most of the color. On three benchmarks this yields models with, on average, 1.79x fewer Gaussians than MaskGS and 5.63x fewer than 3DGS, for PSNR drops of about 0.50 dB and 0.40 dB respectively. The practical upshot is significantly smaller, faster, more memory-efficient 3D scene representations for real-time applications.

What carries the argument

The central object is the spatial mask F(x), a per-pixel scalar computed inside the CUDA rasterizer. For each pixel x, the contributions of the N(x) Gaussians along that ray are summed as M_i * (1 - alpha_i T_i) and normalized by log(1+N(x)). The factor 1 - alpha_i T_i is the inverse of a Gaussian's effective contribution to the ray: a foreground Gaussian with high opacity and high transmittance has alpha_i T_i close to 1, so it contributes almost nothing to F; an occluded or faint Gaussian has alpha_i T_i small, so it contributes a large penalty. The regularizer is the mean squared value of this spatial mask, L_mask = (1/HW) sum_x F(x)^2. Its gradient, derived in closed form, decomposes int

What would settle it

Compute leave-one-out PSNR impact for every Gaussian in a trained SVR-GS model and check whether the kept Gaussians have higher average impact than the pruned ones; if the gap is small or negative, the visibility-weighted opacity proxy is not the right importance signal. A cheaper check: on a scene with fine, faint structures (e.g., a chain-link fence or thin vegetation), count whether such Gaussians survive pruning; the method's own qualitative results show a slight blur on the rear truck wheel, so the proxy may over-penalize partially occluded detail.

Watch

Extended reading notes

Core claim

The paper's central discovery is that aligning the sparsity penalty with the per-ray photometric accumulation makes aggressive pruning much cheaper. Concretely, it proposes rendering a per-pixel spatial mask F(x) = (1/log(1+N(x))) * sum_i M_i (1 - alpha_i T_i) along each ray, where alpha_i T_i is the visibility-weighted opacity of Gaussian i at pixel x. The mask is large when the ray contains many low-importance Gaussians and small when a few foreground splats explain the color. The training loss becomes L_rgb + lambda_F * (mean over pixels of F^2). A gradient analysis shows this regularizer has a self term that pushes down on faint or occluded Gaussians and an occlusion term that rises when

Load-bearing premise

The load-bearing premise is that a Gaussian's visibility-weighted opacity alpha_i(x) T_i(x) is a faithful measure of how much it matters to image quality, so that penalizing Gaussians with small alpha_i T_i prunes exactly the ones whose removal least hurts the rendering.

Editorial extensions

If this is right

  • On average across Tanks&Temples, Deep Blending, and Mip-NeRF360, SVR-GS uses 1.79x fewer Gaussians than MaskGS and 5.63x fewer than 3DGS, with mean PSNR drops of 0.50 dB and 0.40 dB respectively.
  • Because fewer Gaussians survive, on-disk model size, GPU memory during rendering, and per-frame compositing cost all shrink, which is the stated motivation for real-time robotics and AR/VR deployment.
  • The gradient decomposition predicts that the regularizer adapts to the density of active masks: it pushes harder when many masks are on (p≈0.8) and gently when few are on (p≈0.1), so it does not over-prune sparse regions.
  • The design-space ablation on Tanks&Temples shows the chosen forward aggregation (Eq. 4) beats two alternatives (inverse-importance weighting and cumulative-transmittance masking) on both final Gaussian count and reconstruction quality, reaching 0.272M Gaussians at 23.25 dB PSNR.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural extension is to make the spatial mask operate in a multi-scale or semantic-aware way: the visibility-weighted transparency signal could be integrated over patches or over time (for video) to prefer pruning Gaussians that are consistently occluded or faint, not just along single rays.
  • The same importance proxy alpha_i T_i could be recycled for other decisions beyond pruning, such as bit allocation in compression, level-of-detail selection, or scheduling which splats to render first, since it already ranks Gaussians by their contribution to the final image.
  • If the proxy is right, one would predict that the pruned set closely matches the set of Gaussians with the lowest leave-one-out impact on PSNR; a direct comparison would test the core claim more sharply than aggregate metrics, which the paper does not report.
  • The method's fixed lambda_F sweep suggests a controller could adapt the regularization strength during training to hit a target Gaussian budget, rather than selecting lambda_F per scene; the paper uses the same hyperparameters across all datasets, implying stability across scenes without tuning.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes SVR-GS, a spatially variant regularizer for probabilistic mask pruning in 3D Gaussian Splatting. Instead of MaskGS's global mean-mask penalty, SVR-GS renders a per-pixel spatial mask from each Gaussian's visibility-weighted contribution α_i(x)T_i(x), normalizes by 1/log(1+N(x)), and penalizes its squared mean per view. The authors derive the gradient of this mask with respect to per-Gaussian mask probabilities, implement three forward aggregation designs in CUDA, and report experiments on Tanks&Temples, Deep Blending, and Mip-NeRF360. The headline claim is that SVR-GS reduces Gaussian counts by 1.79× vs. MaskGS and 5.63× vs. 3DGS on average, with PSNR drops of 0.50 dB and 0.40 dB, respectively.

Significance. If the claims are supported, the contribution is useful: a principled, locally adaptive pruning signal for probabilistic masks in 3DGS could reduce model size and inference cost with modest quality loss. The paper's strengths include a concrete CUDA implementation, an explicit gradient decomposition in Eq. (9), and a three-way ablation of forward mask designs (Section III.G). The method is clearly positioned against MaskGS and other pruning baselines. However, the empirical support is currently weakened by internally inconsistent headline numbers, lack of uncertainty/error analysis, a possibly overfit hyperparameter selection, and the absence of control experiments that isolate the benefit of the spatial-variant term from the effect of more aggressive pruning. These issues are load-bearing for the central quantitative claims.

major comments (5)
  1. [Abstract, Section III.D, Table I, Conclusions] The paper repeatedly states that the PSNR drop is 'at most 0.5 dB' or 'only 0.50 and 0.40 dB'. Table I shows a 0.74 dB drop on Mip-NeRF360 (27.49 vs. 26.75), which directly contradicts this claim. The authors should correct the abstract and conclusions and qualify the quality cost per dataset; as written, the central summary is factually wrong.
  2. [Table I, Section III.D] Each dataset row in Table I is a single aggregate number, but Tanks&Temples and Mip-NeRF360 comprise multiple scenes with heterogeneous content. No per-scene results, standard deviations, or significance tests are reported. Given that pruning outcomes are highly scene-dependent, the reader cannot assess whether the observed differences are meaningful or noise. Please provide per-scene tables and variance estimates for all metrics and #GS.
  3. [Section III.F, Eq. (11), Table I] The hyperparameter λ_F is swept over the evaluation datasets themselves (λ_F ∈ {1, 1.25, 1.6, 2} × 10⁻⁴), and the main manuscript does not state which λ_F value produced Table I. If λ_F = 1 × 10⁻⁴ was selected for best PSNR, the reported pruning factor may be smaller; if λ_F = 1.25 × 10⁻⁴ was used, the quality drop is larger than the text implies. The trade-off claim is therefore not a fair, held-out evaluation. Please specify the final λ_F and validate it on a separate split or with nested cross-validation.
  4. [Section II.B, Section II.C, Table I] The central claim is that the spatially variant regularizer selectively prunes low-importance Gaussians. However, the paper provides no control experiment that isolates the spatial mechanism from the overall pruning strength. SVR-GS prunes more aggressively than MaskGS on every dataset and is worse on every quality metric. To support the claimed advantage, compare against MaskGS with its λ_m increased to match SVR-GS's final #GS, and against a random-pruning baseline at the same #GS. Without such controls, the 1.79× reduction may reflect only the larger penalty, not the spatial location of the pruning signal; the α_i T_i importance proxy in Eq. (4) is asserted, not validated.
  5. [Abstract, Section I, Section III] The paper claims 'significantly smaller, faster, and more memory-efficient models' but reports only #GS. No measured FPS, training time, peak GPU memory, or on-disk model size is provided. Since the method adds a CUDA forward/backward path for the spatial mask, the computational overhead is not a priori negligible. Please report these quantities or soften the efficiency claims.
minor comments (5)
  1. [Section II.A] Typo: 'A single global average' should be lowercase; 'Gausians' appears elsewhere as 'Gaussians'.
  2. [Eq. (9), Section II.C] The denominator 1 − α_i M_i can vanish if α_i = 1 and M_i = 1. Please discuss the numerical safeguard or clamping used in the CUDA implementation, since this affects gradient stability.
  3. [Eq. (4), Section II.B] The normalization 1/log(1+N(x)) is introduced without derivation or ablation. A brief justification or a sensitivity analysis of this factor would help readers assess whether it is essential to the reported behavior.
  4. [Section III.F] The text says λ_F = 1.25×10⁻⁴ is about 0.3–0.5 dB lower than λ_F = 1×10⁻⁴, yet Table I shows a 0.74 dB drop for Mip-NeRF360. This discrepancy should be reconciled or explicitly attributed to dataset/scene differences.
  5. [Section III.C] The statement 'We use the same hyperparameters across all datasets for fairness' is in tension with the λ_F sweep in Section III.F; please clarify how the final hyperparameter was selected and whether it was frozen before seeing the eval sets.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the spatial-mask regularizer is an added objective and the reported #GS/PSNR numbers are empirical outcomes, not quantities forced by construction.

full rationale

The paper's core proposal is an additional per-pixel regularizer L_mask = (1/HW) Σ F(x)^2, where F(x) = (1/log(1+N(x))) Σ M_i (1 − α_i T_i) (Eqs. 4 and 10), added to the RGB reconstruction loss (Eq. 11). The 'low-importance' label is defined in terms of α_i T_i, but the reported reductions in Gaussian count and PSNR are results of optimizing this objective on standard external benchmarks (Tanks&Temples, Deep Blending, Mip-NeRF360), not quantities derived from the regularizer by construction. There is no fitted parameter that is renamed as a prediction, no load-bearing self-citation, no imported uniqueness theorem, and no ansatz smuggled in via citation. The only mild methodological concern is that λ_F is selected by sweeping on the evaluation datasets (Section III.F), which could inflate the reported trade-off; however, this is a hyperparameter-selection issue, not circularity. The link between the α_i T_i proxy and true reconstruction importance is a heuristic assumption whose failure would be a correctness problem, not a circular derivation.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The method has one free hyperparameter (lambda_F) and relies on standard 3DGS rendering assumptions plus the Gumbel-Softmax mechanism from MaskGS. No new physical entities are introduced. The log-normalization is an ad hoc design choice that affects the loss scale.

free parameters (1)
  • lambda_F = 1e-4 (likely, used in main experiments)
    Regularization strength in Eq. 11. The paper sweeps lambda_F in {1,1.25,1.6,2} x 1e-4 and reports that 1e-4 gives the best PSNR, but does not explicitly state which value produced Table I.
assumptions (3)
  • domain assumption The 3DGS alpha compositing model (Eq. 1-2) accurately describes how Gaussians contribute to a pixel.
    Used as the base rendering model; standard in the 3DGS literature.
  • domain assumption Gumbel-Softmax with straight-through estimation provides valid gradients for the binary masks M_i.
    Inherited from MaskGS; the paper uses the same mask sampling scheme but does not re-derive or justify it.
  • ad hoc to paper The normalization 1/log(1+N(x)) in Eq. (4) makes the spatial mask comparable across rays of different lengths.
    This choice is motivated by intuition but is not derived from a principled criterion; if it distorts the loss, the spatial mask could mis-prioritize rays with many Gaussians.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SVR-GS: Spatially Variant Regularization for Probabilistic Masks in 3D Gaussian Splatting." pith.science (2026). https://pith.science/paper/MGB6UWT7

@misc{pith2026250911116,
  author       = {Pith},
  title        = {Pith review of: SVR-GS: Spatially Variant Regularization for Probabilistic Masks in 3D Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MGB6UWT7}},
  note         = {Machine review of arXiv:2509.11116}
}
abstract

3D Gaussian Splatting (3DGS) enables fast, high-quality novel view synthesis but typically relies on densification followed by pruning to optimize the number of Gaussians. Existing mask-based pruning, such as MaskGS, regularizes the global mean of the mask, which is misaligned with the local per-pixel (per-ray) reconstruction loss that determines image quality along individual camera rays. This paper introduces SVR-GS, a spatially variant regularizer that renders a per-pixel spatial mask from each Gaussian's effective contribution along the ray, thereby applying sparsity pressure where it matters: on low-importance Gaussians. We explore three spatial-mask aggregation strategies, implement them in CUDA, and conduct a gradient analysis to motivate our final design. Extensive experiments on Tanks\&Temples, Deep Blending, and Mip-NeRF360 datasets demonstrate that, on average across the three datasets, the proposed SVR-GS reduces the number of Gaussians by 1.79\(\times\) compared to MaskGS and 5.63\(\times\) compared to 3DGS, while incurring only 0.50 dB and 0.40 dB PSNR drops, respectively. These gains translate into significantly smaller, faster, and more memory-efficient models, making them well-suited for real-time applications such as robotics, AR/VR, and mobile perception.

Figures

Figures reproduced from arXiv: 2509.11116 by the authors.

Figure 1
Figure 1. MaskGS learns per-Gaussian existence probabilities and samples a [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Pipeline overview. (a) Starting from a set of 3D Gaussians Gi , Rasterizer cull splats outside the camera frustum and sort the remainder front￾to-back. The renderer then branches into: (i) an RGB Rendering path that composites colors using Eq. (2); and (ii) a Spatial Mask rendering path that renders a Spatial Mask from the masks Mi (Eq. (4)). The training objective combines the image loss Lrgb with the mask loss as … view at source ↗
Figure 3
Figure 3. C. Backward: Gradient of the Mask Regularizer For a given pixel x, we order the Gaussians along the ray by depth: i = 0 is nearest, then i = 1,...,N−1. We differentiate F with respect to a fixed mask Mi . We use j as a running index in sums and k as a dummy index in products. For clarity of presentation, we omit the pixel index x. We write the transmittance as Tj = ∏ k<j [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (4 more)
Figure 3
Figure 3. Figure 3: Spatial mask F(x) rendered via Eq. (4) (left) alongside the RGB image (right). High intensities correspond to small per-Gaussian contributions αi(x)Ti(x), highlighting occluded or faint structures. By the chain rule, ∂ fj ∂Mi = ∂ fj ∂Tj ∂Tj ∂Mi = (−Mjαj)  −αi Tj 1−αiM…
Figure 4
Figure 4. Figure 4: Qualitative comparison of our method with 3DGS and MaskGS. For each scene, we report PSNR (dB, [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Effect of the forward function on the spatial mask. We visualize the spatial mask F(x) produced by three forward designs; brighter intensities indicate higher penalty / lower importance. (a) Scenario A (Eq. (13)): the inverse-importance weights wi = 1/(αiTi +ε) explode…
Figure 6
Figure 6. Figure 6: Effect of λF on number of Gaussians (top) and quality (bottom). Larger λF prunes more and ends with fewer Gaussians, but PSNR drops. Scenario A falls short because the inverse-importance weights wi = 1/(αiTi +ε) explode when αiTi is tiny, saturat￾ing FA and tagging bro…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 5 linked inside Pith

  1. [1]

    Triangle splatting for real-time radiance field rendering,

    J. Held, R. Vandeghen, A. Deliege, A. Hamdi, S. Giancola, A. Cioppa, A. Vedaldi, B. Ghanem, A. Tagliasacchi, and M. Van Droogenbroeck, “Triangle splatting for real-time radiance field rendering,”arXiv preprint arXiv:2505.19175, 2025

  2. [2]

    A survey on 3d gaussian splatting applications: Segmentation, editing, and generation,

    S. He, P. Ji, Y . Yang, C. Wang, J. Ji, Y . Wang, and H. Ding, “A survey on 3d gaussian splatting applications: Segmentation, editing, and generation,”arXiv preprint arXiv:2508.09977, 2025

  3. [3]

    3d convex splatting: Radiance field rendering with 3d smooth convexes,

    J. Held, R. Vandeghen, A. Hamdi, A. Deliege, A. Cioppa, S. Giancola, A. Vedaldi, B. Ghanem, and M. Van Droogenbroeck, “3d convex splatting: Radiance field rendering with 3d smooth convexes,” inPro- ceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 21 360–21 369

  4. [4]

    Linprim: Linear primitives for differentiable volumetric rendering,

    N. von L ¨utzow and M. Nießner, “Linprim: Linear primitives for differentiable volumetric rendering,”arXiv preprint arXiv:2501.16312, 2025

  5. [5]

    Arbitrary-scale 3d gaussian super- resolution,

    H. Zeng, Y . Bai, and Y . Fu, “Arbitrary-scale 3d gaussian super- resolution,”arXiv preprint arXiv:2508.16467, 2025

  6. [6]

    Nerf in the palm of your hand: Corrective augmentation for robotics via novel-view synthesis,

    A. Zhou, M. J. Kim, L. Wang, P. Florence, and C. Finn, “Nerf in the palm of your hand: Corrective augmentation for robotics via novel-view synthesis,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 17 907–17 917

  7. [7]

    Gs-planner: A gaussian-splatting-based planning framework for active high-fidelity reconstruction,

    R. Jin, Y . Gao, Y . Wang, Y . Wu, H. Lu, C. Xu, and F. Gao, “Gs-planner: A gaussian-splatting-based planning framework for active high-fidelity reconstruction,” in2024 IEEE/RSJ International Conference on Intel- ligent Robots and Systems (IROS), 2024, pp. 11 202–11 209

  8. [8]

    Hgs-mapping: Online dense mapping using hybrid gaus- sian representation in urban scenes,

    K. Wu, K. Zhang, Z. Zhang, M. Tie, S. Yuan, J. Zhao, Z. Gan, and W. Ding, “Hgs-mapping: Online dense mapping using hybrid gaus- sian representation in urban scenes,”IEEE Robotics and Automation Letters, vol. 9, no. 11, pp. 9573–9580, 2024

Show all 45 references
  1. [9]

    Gaussiangrasper: 3d language gaussian splatting for open-vocabulary robotic grasping,

    Y . Zheng, X. Chen, Y . Zheng, S. Gu, R. Yang, B. Jin, P. Li, C. Zhong, Z. Wang, L. Liu, C. Yang, D. Wang, Z. Chen, X. Long, and M. Wang, “Gaussiangrasper: 3d language gaussian splatting for open-vocabulary robotic grasping,”IEEE Robotics and Automation Letters, vol. 9, no. 9,...

  2. [10]

    Egolifter: Open-world 3d segmentation for egocentric perception,

    Q. Gu, Z. Lv, D. Frost, S. Green, J. Straub, and C. Sweeney, “Egolifter: Open-world 3d segmentation for egocentric perception,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 382–400

  3. [11]

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

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoor- thi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,”Communications of the ACM, vol. 65, no. 1, pp. 99–106, 2021

  4. [12]

    Frugalnerf: Fast convergence for extreme few-shot novel view synthesis without learned priors,

    C.-Y . Lin, C.-H. Wu, C.-H. Yeh, S.-H. Yen, C. Sun, and Y .-L. Liu, “Frugalnerf: Fast convergence for extreme few-shot novel view synthesis without learned priors,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2025, pp. 11 227–11 238

  5. [13]

    E2nerf: Event enhanced neural radiance fields from blurry images,

    Y . Qi, L. Zhu, Y . Zhang, and J. Li, “E2nerf: Event enhanced neural radiance fields from blurry images,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 13 254– 13 264

  6. [14]

    Lush-nerf: Light- ing up and sharpening nerfs for low-light scenes,

    Z. Qu, K. Xu, G. P. Hancke, and R. W. Lau, “Lush-nerf: Light- ing up and sharpening nerfs for low-light scenes,”arXiv preprint arXiv:2411.06757, 2024

  7. [15]

    Object-aware gaussian splatting for robotic manipulation,

    Y . Li and D. Pathak, “Object-aware gaussian splatting for robotic manipulation,” inICRA 2024 Workshop on 3D Visual Representations for Robot Manipulation, 2024. [Online]. Available: https://openreview.net/forum?id=gdRI43hDgo

  8. [16]

    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.”ACM Trans. Graph., vol. 42, no. 4, pp. 139–1, 2023

  9. [17]

    3d gaussian splatting in robotics: A survey,

    S. Zhu, G. Wang, X. Kong, D. Kong, and H. Wang, “3d gaussian splatting in robotics: A survey,”arXiv preprint arXiv:2410.12262, 2024

  10. [18]

    Dreamdrive: Generative 4d scene modeling from street view images,

    J. Mao, B. Li, B. Ivanovic, Y . Chen, Y . Wang, Y . You, C. Xiao, D. Xu, M. Pavone, and Y . Wang, “Dreamdrive: Generative 4d scene modeling from street view images,” in2025 IEEE International Conference on Robotics and Automation (ICRA), 2025, pp. 367–374

  11. [19]

    Spotlesssplats: Ignoring distractors in 3d gaussian splatting,

    S. Sabour, L. Goli, G. Kopanas, M. Matthews, D. Lagun, L. Guibas, A. Jacobson, D. Fleet, and A. Tagliasacchi, “Spotlesssplats: Ignoring distractors in 3d gaussian splatting,”ACM Trans. Graph., vol. 44, no. 2, Apr. 2025. [Online]. Available: https://doi.org/10.1145/3727143

  12. [20]

    Mini-splatting: Representing scenes with a constrained number of gaussians,

    G. Fang and B. Wang, “Mini-splatting: Representing scenes with a constrained number of gaussians,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 165–181

  13. [21]

    Cob-gs: Clear object boundaries in 3dgs segmentation based on boundary-adaptive gaussian splitting,

    J. Zhang, J. Jiang, Y . Chen, K. Jiang, and X. Liu, “Cob-gs: Clear object boundaries in 3dgs segmentation based on boundary-adaptive gaussian splitting,” inCVPR, 2025

  14. [22]

    Enliveninggs: Ac- tive locomotion of 3dgs,

    S. Shen, T. Shao, K. Zhou, C. Jiang, and Y . Yang, “Enliveninggs: Ac- tive locomotion of 3dgs,” in2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 896–905

  15. [23]

    High-fidelity slam using gaussian splatting with rendering-guided densification and regularized optimization,

    S. Sun, M. Mielle, A. J. Lilienthal, and M. Magnusson, “High-fidelity slam using gaussian splatting with rendering-guided densification and regularized optimization,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 10 476– 10 482

  16. [24]

    Gaussianroom: Improving 3d gaussian splatting with sdf guidance and monocular cues for indoor scene reconstruction,

    H. Xiang, X. Li, K. Cheng, X. Lai, W. Zhang, Z. Liao, L. Zeng, and X. Liu, “Gaussianroom: Improving 3d gaussian splatting with sdf guidance and monocular cues for indoor scene reconstruction,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, ...

  17. [25]

    Compressed 3d gaussian splatting for accelerated novel view synthesis,

    S. Niedermayr, J. Stumpfegger, and R. Westermann, “Compressed 3d gaussian splatting for accelerated novel view synthesis,” inProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 10 349–10 358

  18. [26]

    Contextgs: Compact 3d gaussian splatting with anchor level context model,

    Y . Wang, Z. Li, L. Guo, W. Yang, A. Kot, and B. Wen, “Contextgs: Compact 3d gaussian splatting with anchor level context model,” Advances in neural information processing systems, vol. 37, pp. 51 532–51 551, 2024

  19. [27]

    Compact 3d scene representation via self-organizing gaussian grids,

    W. Morgenstern, F. Barthel, A. Hilsmann, and P. Eisert, “Compact 3d scene representation via self-organizing gaussian grids,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 18–34

  20. [28]

    Radsplat: Radiance field-informed gaussian splatting for robust real- time rendering with 900+ fps,

    M. Niemeyer, F. Manhardt, M.-J. Rakotosaona, M. Oechsle, D. Duck- worth, R. Gosula, K. Tateno, J. Bates, D. Kaeser, and F. Tombari, “Radsplat: Radiance field-informed gaussian splatting for robust real- time rendering with 900+ fps,” in2025 International Conference on 3D Visio...

  21. [30]

    Lp-3dgs: Learning to prune 3d gaussian splatting,

    Z. Zhang, T. Song, Y . Lee, L. Yang, C. Peng, R. Chellappa, and D. Fan, “Lp-3dgs: Learning to prune 3d gaussian splatting,”Advances in Neural Information Processing Systems, vol. 37, pp. 122 434–122 457, 2024

  22. [31]

    Compact 3d gaussian representation for radiance field,

    J. C. Lee, D. Rho, X. Sun, J. H. Ko, and E. Park, “Compact 3d gaussian representation for radiance field,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 21 719–21 728

  23. [32]

    End-to-end rate-distortion optimized 3d gaussian representation,

    H. Wang, H. Zhu, T. He, R. Feng, J. Deng, J. Bian, and Z. Chen, “End-to-end rate-distortion optimized 3d gaussian representation,” in European Conference on Computer Vision. Springer, 2024, pp. 76–92

  24. [33]

    Light- gaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps,

    Z. Fan, K. Wang, K. Wen, Z. Zhu, D. Xu, Z. Wanget al., “Light- gaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps,”Advances in neural information processing systems, vol. 37, pp. 140 138–140 158, 2024

  25. [34]

    Efficientgs: Streamlining gaussian splatting for large-scale high-resolution scene representation,

    W. Liu, T. Guan, B. Zhu, L. Xu, Z. Song, D. Li, Y . Wang, and W. Yang, “Efficientgs: Streamlining gaussian splatting for large-scale high-resolution scene representation,”IEEE MultiMedia, 2025

  26. [35]

    Reducing the memory footprint of 3d gaussian splatting,

    P. Papantonakis, G. Kopanas, B. Kerbl, A. Lanvin, and G. Drettakis, “Reducing the memory footprint of 3d gaussian splatting,”Proceed- ings of the ACM on Computer Graphics and Interactive Techniques, vol. 7, no. 1, pp. 1–17, 2024

  27. [36]

    Pup 3d-gs: Principled uncertainty pruning for 3d gaussian splatting,

    A. Hanson, A. Tu, V . Singla, M. Jayawardhana, M. Zwicker, and T. Goldstein, “Pup 3d-gs: Principled uncertainty pruning for 3d gaussian splatting,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 5949–5958

  28. [37]

    Taming 3dgs: High-quality radiance fields with limited resources,

    S. S. Mallick, R. Goel, B. Kerbl, M. Steinberger, F. V . Carrasco, and F. De La Torre, “Taming 3dgs: High-quality radiance fields with limited resources,” inSIGGRAPH Asia 2024 Conference Papers, 2024, pp. 1–11

  29. [38]

    On reducing the number of gaussians for radiance field real-time rendering,

    O. Mahmoud and M. Gendrin, “On reducing the number of gaussians for radiance field real-time rendering,” in2024 Joint International Conference on Digital Arts, Media and Technology with ECTI North- ern Section Conference on Electrical, Electronics, Computer and Telecommunicati...

  30. [39]

    Maskgaussian: Adaptive 3d gaussian representation from probabilistic masks,

    Y . Liu, Z. Zhong, Y . Zhan, S. Xu, and X. Sun, “Maskgaussian: Adaptive 3d gaussian representation from probabilistic masks,” inPro- ceedings of the Computer Vision and Pattern Recognition Conference (CVPR), June 2025, pp. 681–690

  31. [40]

    Categorical reparameterization with gumbel-softmax,

    E. Jang, S. Gu, and B. Poole, “Categorical reparameterization with gumbel-softmax,”arXiv preprint arXiv:1611.01144, 2016

  32. [41]

    Tanks and temples: benchmarking large-scale scene reconstruction,

    A. Knapitsch, J. Park, Q.-Y . Zhou, and V . Koltun, “Tanks and temples: benchmarking large-scale scene reconstruction,”ACM Trans. Graph., vol. 36, no. 4, Jul. 2017. [Online]. Available: https://doi.org/10.1145/3072959.3073599

  33. [42]

    Deep blending for free-viewpoint image-based rendering,

    P. Hedman, J. Philip, T. Price, J.-M. Frahm, G. Drettakis, and G. Brostow, “Deep blending for free-viewpoint image-based rendering,”ACM Trans. Graph., vol. 37, no. 6, Dec. 2018. [Online]. Available: https://doi.org/10.1145/3272127.3275084

  34. [43]

    Mip-nerf 360: Unbounded anti-aliased neural radiance fields,

    J. T. Barron, B. Mildenhall, D. Verbin, P. P. Srinivasan, and P. Hedman, “Mip-nerf 360: Unbounded anti-aliased neural radiance fields,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 5470–5479

  35. [44]

    Image quality metrics: Psnr vs. ssim,

    A. Hor ´e and D. Ziou, “Image quality metrics: Psnr vs. ssim,” in 2010 20th International Conference on Pattern Recognition, 2010, pp. 2366–2369

  36. [45]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. Bovik, H. Sheikh, and E. Simoncelli, “Image quality assessment: from error visibility to structural similarity,”IEEE Trans- actions on Image Processing, vol. 13, no. 4, pp. 600–612, 2004

  37. [46]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,”2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 586–595, 2018. [Online]. Available: https://api.semanticscholar.org/Cor...

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.