Pith. sign in

REVIEW 6 major objections 6 minor 56 references

Wavelet-GS: 3D Gaussian Splatting with Wavelet Decomposition

T0 review · 6 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Paper claims that splitting a 3D point cloud into low- and high-frequency parts with wavelets, then training each part separately inside 3D Gaussian Splatting, yields more complete scene structures and sharper details in novel view…

desk verdict Promising idea, missing core: the 3D wavelet transform on point clouds is never defined, and the SOTA claims outrun the reported numbers. read the letter →

arxiv 2507.12498 v2 pith:5EJHKKCD submitted 2025-07-16 cs.GR

Beizhen Zhao , Yifan Zhou , Sicheng Yu , Zijian Wang , Hao Wang This is my paper
classification cs.GR
keywords 3DreconstructionGaussiansplattingwaveletdecompositionpoint-basedrenderingnovelviewsynthesisfrequencyseparationrelightingsparseviews
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 separating a 3D point cloud into low- and high-frequency parts with a wavelet transform, then training each part with its own strategy inside 3D Gaussian Splatting, produces more complete scene structures and finer detail in novel view synthesis. The low-frequency branch is voxelized and grown/pruned to capture global outlines, while the high-frequency branch is trained to restore edges and textures and is paired with a relighting module that uses 2D wavelet decompositions of training images to model lighting. The paper reports that this combined framework beats prior 3DGS variants across four datasets (Waymo, Mip-NeRF360, Tanks&Temples, JHU-Drone) on PSNR, SSIM, and LPIPS, with the largest gains on sparse-view driving scenes. If correct, it would mean that explicit frequency separation inside a splatting pipeline is a practical route to sharper, more coherent scene reconstruction.

What carries the argument

The central object is the 3D wavelet decomposition of a point cloud, which splits the point set into Plow and Phigh using a scaling function phi and a wavelet function psi along the three spatial axes (Eqs. 9-10), exploiting the linearity of the wavelet transform to recombine them (Eq. 11). Each branch is voxelized and converted into Gaussian parameters by separate MLPs: the low-frequency branch uses a grow-and-prune voxel strategy inspired by Scaffold-GS, while the high-frequency branch uses a deviation thresholding mask to control voxel growth and a relight module (environment map times SH visibility) to compute per-Gaussian color. A 2D wavelet decomposition of the training images supplies structural features via an MLP, and the Laplacian-Wavelet loss LL-W penalizes differences across Laplacian pyramid levels and 2D DWT bands. These pieces together let the pipeline train low- and high-frequency content independently and recombine them at render time.

What would settle it

Run the 3D DWT exactly as specified by Eqs. 9-10 on a real COLMAP point cloud from one of the four datasets and check whether Plow and Phigh are finite point sets with well-defined coordinates that voxelize cleanly; if the transform requires a regular grid that the paper never defines, the decomposition is ill-posed and the method's core step does not exist as stated.

Watch

Extended reading notes

Core claim

Wavelet-GS integrates a 3D discrete wavelet transform into 3D Gaussian Splatting, decomposing the input point cloud P into a low-frequency component Plow and a high-frequency component Phigh along the XYZ axes (Eqs. 8-10). Plow is voxelized into Vlow and rendered by neural Gaussians that are grown and pruned by gradient and opacity, capturing the global structural skeleton; Phigh is voxelized into Vhigh and rendered by a separate set of neural Gaussians whose colors are modulated by a relight module, restoring fine edges and textures. A 2D wavelet transform on the training images produces structural features that drive the relight module via spherical harmonics, and a Laplacian-Wavelet loss supervises multi-scale detail. The two branches are fused by the linearity of the wavelet transform (G = Gl + Gh), and the authors report that this decoupled, frequency-aware optimization surpasses prior state-of-the-art 3DGS variants on all four tested datasets.

Load-bearing premise

The load-bearing premise is that a discrete wavelet transform can be meaningfully applied to a sparse, irregular 3D point cloud along the XYZ axes so that the outputs Plow and Phigh are still point clouds that can be voxelized into Gaussians, but the paper does not specify the underlying grid, the decomposition level, or how wavelet coefficients map back to 3D positions.

Editorial extensions

If this is right

  • If the reported gains hold, frequency-separated optimization becomes a viable template for other point-based renderers, not just 3DGS.
  • The 2D-wavelet-driven relight module could transfer to any neural renderer that needs lighting robustness under sparse views.
  • The Laplacian-Wavelet loss could be reused as a general multi-scale detail loss in image- and radiance-field training.
  • On sparse-view datasets like Waymo, the method claims the largest improvements, suggesting frequency decomposition helps most when supervision is scarce.
  • The reported state-of-the-art numbers on four datasets imply that wavelet decomposition, not just extra capacity, is what drives the improvement.

Reading between the lines

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

  • The 3D DWT step as written is a voxel-grid convolution in disguise: the point cloud must first be sampled onto a regular grid for Eqs. 9-10 to apply, so the 'wavelet' part may be equivalent to a multiresolution voxel feature split rather than a true point-cloud transform.
  • The paper's ablation shows removing the 3D wavelet costs about 0.76 dB PSNR on Waymo while removing the 2D wavelet costs 0.43 dB and removing the individual strategy costs 1.12 dB; the largest single contributor may be the training strategy, not the wavelet decomposition itself.
  • A cheap test of the wavelet family's importance would be to replace coif1 with a simple high-pass/low-pass filter pair on the voxel grid; if the gains persist, the specific wavelet basis is not the active ingredient.
  • The method's dependence on voxelization suggests it could combine naturally with octree or hierarchical Gaussian structures to scale to unbounded scenes.
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

6 major / 6 minor

Summary. The paper proposes Wavelet-GS, a 3D Gaussian Splatting framework that applies 3D wavelet decomposition to the input point cloud, splitting it into low- and high-frequency branches that are trained with separate strategies, and additionally applies a 2D wavelet transform to training images to guide a relight module. The low-frequency branch is voxelized and trained with a grow-and-prune strategy, while the high-frequency branch is trained with a Laplacian-Wavelet loss and a LumiGauss-style relight module. Experiments on Waymo, JHU-Drone, Tanks&Temples, and Mip-NeRF360 report strong quantitative results, and ablations on the Waymo dataset are used to justify each component.

Significance. If the method were fully specified and reproducible, a principled frequency-separated Gaussian training scheme would be a useful contribution to the 3DGS literature, and the inclusion of four datasets plus ablations over wavelet families is a positive feature. However, the central construction—the 3D wavelet decomposition of an unorganized point cloud—is never concretely defined, the claimed fusion of the two branches is not connected to the rendering equation, and the paper's own Table 1 contradicts the headline claim of surpassing all methods on all four datasets. As submitted, the significance of the empirical results cannot be evaluated because the method is not implementable from the text.

major comments (6)
  1. [Section 3.3, Eqs. (8)-(11)] The 3D DWT is applied to 'the 3D point cloud P', but P is a set of unorganized points; Eqs. (9)-(10) treat P(m,n,p) as values on a regular grid indexed by (m,n,p) and evaluate scaling/wavelet functions at grid indices, without specifying the grid resolution, extents, decomposition level, boundary handling, or the mapping from wavelet coefficients back to 3D point coordinates. A standard single-level 3D DWT of a voxel grid yields eight subbands (LLL, LLH, LHL, LHH, HLL, HLH, HHL, HHH), not two point clouds, and Eq. (11) is not the inverse-DWT reconstruction formula: synthesis filters and upsampling are required. As written, the central operation is unimplementable, so the low/high branch separation and all subsequent results cannot be attributed to the described mechanism.
  2. [Section 3.5, Eq. (16)] The fusion rule G(x)=Gl(x)+Gh(x) is asserted to follow from wavelet linearity, but the paper never defines the addition of two sets of anisotropic Gaussians or how the two branches are combined in the rendering pass; the rendering equation (7) is the standard alpha-blending formula with a single set of Gaussians. The manuscript needs an explicit forward pass showing how Plow and Phigh become Gaussian parameters and how their rendered images are fused.
  3. [Section 4.1 and Section 3.5] Key implementation details are missing: no voxelization threshold or resolution for Vlow/Vhigh beyond the global voxel size 0.001, no coordinate normalization for unbounded scenes, no definition of the high-frequency deviation mask thresholds, no grow/prune schedules, no optimizer settings, and no code or supplementary implementation. Without these, the experiments are not reproducible and the 'individual optimization strategy' cannot be evaluated.
  4. [Table 1, Tanks&Temples row] The abstract and introduction claim that the framework 'surpasses existing state-of-the-art results on all four 3D datasets', but on Tanks&Temples the reported PSNR (24.40) is lower than Octree-GS (24.54); the same table also shows equal SSIM (0.863) with Octree-GS on that dataset. The claim needs to be restated per metric or the results corrected, and the conclusion that the method surpasses all methods on all four datasets is not supported by the presented numbers.
  5. [Section 3.4, Eq. (15)] The definition LSH = E[min(0, Li(M))^2] is confusing: min(0, ·) is non-positive, and squaring it gives a positive quantity but this is not the standard squared-SH loss convention (which uses the positive part of the irradiance). Please clarify the sign convention and justify the choice; as written, the loss can encourage negative radiance values and its optimization target is unclear.
  6. [Section 3.5.2, Eqs. (23)-(24)] The relight color formulas have dimension mismatches: in Eq. (23), n_k is a normal vector but is used as n_k^T M_env n_k where M_env is not defined as a matrix; in Eq. (24), rho_k in R^3 is multiplied by an SH sum whose coefficient indexing is inconsistent with Eq. (14). Please provide a precise tensor specification for the relight module.
minor comments (6)
  1. [Figure 2 caption] The caption contains typos: 'gradint' should be 'gradient', 'compoment' should be 'component', and 'in the same time' should be 'at the same time'.
  2. [Section 3.4] The sentence 'we follow [13] and use LSH-env:' is incomplete; Eq. (15) follows, but the relation of LSH to the SH environment loss and to the preceding SH expansion should be stated explicitly.
  3. [Section 4.3, Table 3] Table 3 is described as comparing wavelet families on 'real-world datasets', but it reports results only on Waymo; the caption and text should specify the dataset.
  4. [Abstract and Section 3.4] The phrase 'simulating radiance variations' is vague; the 2D wavelet decomposition is applied to training images, but the connection between the wavelet coefficients, the structural feature M in Eq. (13), and the final radiance is not made precise.
  5. [Throughout] The paper repeatedly uses 'function' where 'module' or 'component' is meant (e.g., '3D wavelet decomposition function', 'loss function LL-W'); please standardize terminology.
  6. [Eq. (22)] The symbol L is used both for the number of Laplacian pyramid levels and for the pyramid representation L1^(l); please disambiguate the notation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the pipeline is a supervised reconstruction and the wavelet steps are preprocessing, not fitted predictions.

full rationale

The paper is a supervised reconstruction method: 3DGS Gaussians are optimized against photometric losses (Eqs. 7, 19, 25) on external benchmarks (Mip-NeRF360, Waymo, Tanks&Temples, JHU-Drone). The claimed 'prediction' is novel-view rendering from trained Gaussians, which is not equivalent to any fitted input by construction. The 3D wavelet decomposition (Eqs. 8-11) is introduced as a preprocessing or initialization step; whether it is mathematically well-defined for a sparse point cloud is a reproducibility and correctness concern, not a circularity, because the text never claims that state-of-the-art performance is logically entailed by the wavelet identities. The relight module is explicitly ported from LumiGauss [13], an external citation, and the two same-group citations ([19] dataset, [35] GVKF) are used for evaluation and a background limitation remark, not to justify the core derivation. Equation 16 (G = G_l + G_h) is a fusion formulation; even if it is not actually used in the rendering equation, that is a gap in the write-up, not a fit disguised as a prediction. No step reduces the output to its inputs by definition, so no circularity is found.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The method's central operation is under-specified; the paper hides the implementation of the point-cloud DWT behind equations that apply only to regular grids, and reports no sensitivity analysis for the many fixed hyperparameters (voxel size, loss weights, growth/pruning thresholds). These choices, not derived from theory, are effectively free parameters.

free parameters (6)
  • voxel_size = 0.001
    Set for all scenes without units or sensitivity analysis; controls the granularity of voxelization that generates Gaussians (Section 4.1).
  • neighbor_count_k = 10
    Number of neighbor nodes used in the voxel graph (Section 4.1); fixed across all datasets.
  • loss_weights = lambda_SSIM=0.2, lambda_vol=0.01, lambda_SH=0.05, lambda_L-W unspecified
    Ad hoc weights for Lpixel, LSH and LL-W; lambda_L-W in Eq. 25 is not given a value.
  • wavelet_family = coif1
    Selected by ablation on the Waymo dataset (Table 3), then used for all reported results; selection on the evaluation set is a form of tuning.
  • grow_prune_thresholds = not specified
    Voxel growth uses average gradient plus a random picking strategy; pruning uses an opacity threshold described only as 'sufficient opacity' (Section 3.5.1).
  • high_frequency_mask_thresholds = not specified
    Masking of 'excessively large or small' high-frequency deviations is not quantified (Section 3.5.2).
assumptions (4)
  • standard math Wavelet transform linearity: T{f1+f2} = T{f1} + T{f2}
    Invoked in Eq. 3 to justify fusing low- and high-frequency Gaussians in Eq. 16; mathematically true for grid signals but does not extend to Gaussian parameter spaces.
  • domain assumption Point cloud can be decomposed by 3D DWT along XYZ
    Section 3.3, Eqs. 8-11 assume P(m,n,p) exists on a regular grid, which is not true for sparse point clouds; no construction is provided.
  • domain assumption SH-based radiance transfer and LumiGauss two-stage relight training apply to the high-frequency component
    Section 3.4 and 3.5.2 port the env-map SH model and loss from LumiGauss [13] without re-derivation for the wavelet-separated branch.
  • ad hoc to paper G(x) = G_l(x) + G_h(x) via wavelet linearity
    Eq. 16 asserts additive fusion of Gaussian parameters based on wavelet linearity, which does not follow for the parameters (position, scale, opacity, color) of Gaussians.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Wavelet-GS: 3D Gaussian Splatting with Wavelet Decomposition." pith.science (2026). https://pith.science/paper/5EJHKKCD

@misc{pith2026250712498,
  author       = {Pith},
  title        = {Pith review of: Wavelet-GS: 3D Gaussian Splatting with Wavelet Decomposition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5EJHKKCD}},
  note         = {Machine review of arXiv:2507.12498}
}
read the original abstract

3D Gaussian Splatting (3DGS) has revolutionized 3D scene reconstruction, which effectively balances rendering quality, efficiency, and speed. However, existing 3DGS approaches usually generate plausible outputs and face significant challenges in complex scene reconstruction, manifesting as incomplete holistic structural outlines and unclear local lighting effects. To address these issues simultaneously, we propose a novel decoupled optimization framework, which integrates wavelet decomposition into 3D Gaussian Splatting and 2D sampling. Technically, through 3D wavelet decomposition, our approach divides point clouds into high-frequency and low-frequency components, enabling targeted optimization for each. The low-frequency component captures global structural outlines and manages the distribution of Gaussians through voxelization. In contrast, the high-frequency component restores intricate geometric and textural details while incorporating a relight module to mitigate lighting artifacts and enhance photorealistic rendering. Additionally, a 2D wavelet decomposition is applied to the training images, simulating radiance variations. This provides critical guidance for high-frequency detail reconstruction, ensuring seamless integration of details with the global structure. Extensive experiments on challenging datasets demonstrate our method achieves state-of-the-art performance across various metrics, surpassing existing approaches and advancing the field of 3D scene reconstruction.

Figures

Figures reproduced from arXiv: 2507.12498 by the authors.

Figure 1
Figure 1. Overview of our model. Through 3D wavelet decomposition, the 3D point cloud is divided into low frequency and high [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Framework of Wavelet-GS. We begin by preprocessing the point cloud by voxelization and 3D wavelet decomposition. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. 3D wavelet decomposition. The input data for the [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Simulation for the structural feature We use 2D [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Comparison Results. Visual differences are highlighted with red insets for better clarity. Our approach consistently [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 37 canonical work pages

  1. [1]

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. 2022. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 5470–5479

    2022

  2. [2]

    Debra Charnley and Rod Blissett. 1989. Surface reconstruction from outdoor image sequences. Image and Vision Computing 7, 1 (1989), 10–16

    work page 1989

  3. [3]

    Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. 2024. PGSR: Planar-based Gaussian Splatting for Efficient and High-Fidelity Surface Reconstruction. arXiv preprint arXiv:2406.06521 (2024)

    arXiv 2024

  4. [4]

    Guikun Chen and Wenguan Wang. 2024. A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890 (2024)

    arXiv 2024

  5. [5]

    Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. 2022. Depth- supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 12882–12891

    work page 2022

  6. [6]

    Kyle Gao, Yina Gao, Hongjie He, Dening Lu, Linlin Xu, and Jonathan Li. 2022. Nerf: Neural radiance field in 3d vision, a comprehensive review. arXiv preprint arXiv:2210.00379 (2022)

    arXiv 2022

  7. [7]

    Amara Graps. 1995. An introduction to wavelets. IEEE computational science and engineering 2, 2 (1995), 50–61

    work page 1995

  8. [8]

    Robin Green. 2003. Spherical harmonic lighting: The gritty details. In Archives of the game developers conference , Vol. 56. 4

    work page 2003

Show all 56 references
  1. [9]

    Martin Habbecke and Leif Kobbelt. 2007. A surface-growing approach to multi- view stereo reconstruction. In 2007 IEEE Conference on Computer Vision and Pattern Recognition. IEEE, 1–8

    2007

  2. [10]

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao

  3. [11]

    Ying Jiang, Chang Yu, Tianyi Xie, Xuan Li, Yutao Feng, Huamin Wang, Minchen Li, Henry Lau, Feng Gao, Yin Yang, et al. 2024. Vr-gs: A physical dynamics-aware interactive gaussian splatting system in virtual reality. In ACM SIGGRAPH 2024 Conference Papers. 1–1

    2024

  4. [12]

    Hailin Jin, Stefano Soatto, and Anthony J Yezzi. 2005. Multi-view stereo re- construction of dense shape and complex appearance. International Journal of Computer Vision 63 (2005), 175–189

    2005

  5. [13]

    Joanna Kaleta, Kacper Kania, Tomasz Trzcinski, and Marek Kowalski. 2024. Lu- miGauss: High-Fidelity Outdoor Relighting with 2D Gaussian Splatting. arXiv preprint arXiv:2408.04474 (2024)

    2024 arXiv

  6. [14]

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis

  7. [15]

    Bernhard Kerbl, Andreas Meuleman, Georgios Kopanas, Michael Wimmer, Alexandre Lanvin, and George Drettakis. 2024. A hierarchical 3d gaussian repre- sentation for real-time rendering of very large datasets. ACM Transactions on Graphics (TOG) 43, 4 (2024), 1–15

    2024

  8. [16]

    Hansung Kim, Jean-Yves Guillemaut, Takeshi Takai, Muhammad Sarim, and Adrian Hilton. 2012. Outdoor dynamic 3-D scene reconstruction. IEEE Transac- tions on Circuits and Systems for Video Technology 22, 11 (2012), 1611–1622

    2012

  9. [17]

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. 2017. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG) 36, 4 (2017), 1–13

    2017

  10. [18]

    Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt

  11. [19]

    Xijun Liu, Yifan Zhou, Yuxiang Guo, Rama Chellappa, and Cheng Peng. 2024. An Immersive Multi-Elevation Multi-Seasonal Dataset for 3D Reconstruction and Visualization. arXiv preprint arXiv:2412.14418 (2024)

    2024 arXiv

  12. [20]

    Yang Liu, Chuanchen Luo, Lue Fan, Naiyan Wang, Junran Peng, and Zhaoxiang Zhang. 2025. Citygaussian: Real-time high-quality large-scale scene rendering with gaussians. In European Conference on Computer Vision . Springer, 265–282

    2025

  13. [21]

    Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. 2021. Mixture of volumetric primitives for efficient neural rendering. ACM Transactions on Graphics (ToG) 40, 4 (2021), 1–13

    2021

  14. [22]

    Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. 2024. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 20654–20664

    2024

  15. [23]

    He Lyu, Ningyu Sha, Shuyang Qin, Ming Yan, Yuying Xie, and Rongrong Wang

  16. [24]

    Zhiliang Ma and Shilong Liu. 2018. A review of 3D reconstruction techniques in civil engineering and their applications. Advanced Engineering Informatics 37 (2018), 163–174

    2018

  17. [25]

    Stephane Mallat. 1999. A wavelet tour of signal processing

    1999

  18. [26]

    Stephane G Mallat. 1989. A theory for multiresolution signal decomposition: the wavelet representation. IEEE transactions on pattern analysis and machine intelligence 11, 7 (1989), 674–693

    1989

  19. [27]

    Yves Meyer. 1992. Wavelets and operators: volume 1 . Number 37. Cambridge university press

    1992

  20. [28]

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. 2021. Nerf: Representing scenes as neural radiance fields for view synthesis. Commun. ACM 65, 1 (2021), 99–106

    2021

  21. [29]

    Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Giebenhain, and Matthias Nießner. 2024. Gaussianavatars: Photorealistic head avatars with rigged 3d gaussians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 20299–20309

    2024

  22. [30]

    Ravi Ramamoorthi and Pat Hanrahan. 2001. An efficient representation for irradiance environment maps. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques . 497–500

    2001

  23. [31]

    Kerui Ren, Lihan Jiang, Tao Lu, Mulin Yu, Linning Xu, Zhangkai Ni, and Bo Dai

  24. [32]

    Shunsuke Saito, Gabriel Schwartz, Tomas Simon, Junxuan Li, and Giljoo Nam

  25. [33]

    Thomas Schöps, Torsten Sattler, Christian Häne, and Marc Pollefeys. 2017. Large- scale outdoor 3D reconstruction on a mobile device. Computer Vision and Image Understanding 157 (2017), 151–166

    2017

  26. [34]

    Steven M Seitz, Brian Curless, James Diebel, Daniel Scharstein, and Richard Szeliski. 2006. A comparison and evaluation of multi-view stereo reconstruction algorithms. In 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR’06), Vol. 1. IEEE, 519–528

    2006

  27. [35]

    Gaochao Song, Chong Cheng, and Hao Wang. 2024. GVKF: Gaussian Voxel Kernel Functions for Highly Efficient Surface Reconstruction in Open Scenes. arXiv preprint arXiv:2411.01853 (2024)

    2024 arXiv

  28. [36]

    arXiv preprint arXiv:2403.17898 (2024)

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

    2024 arXiv

  29. [37]

    Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al

  30. [38]

    In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition

    Relightable gaussian codec avatars. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition . 130–141

  31. [39]

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. 2004. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing 13, 4 (2004), 600–612

    2004

  32. [40]

    Tong Wu, Yu-Jie Yuan, Ling-Xiao Zhang, Jie Yang, Yan-Pei Cao, Ling-Qi Yan, and Lin Gao. 2024. Recent advances in 3d gaussian splatting. Computational Visual Media 10, 4 (2024), 613–642

    2024

  33. [41]

    Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chen- fanfu Jiang. 2024. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 4389–4398

    2024

  34. [42]

    Cheng Sun, Min Sun, and Hwann-Tzong Chen. 2022. Direct voxel grid optimiza- tion: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 5459–5469

    2022

  35. [43]

    Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. 2024. Street gaussians: Model- ing dynamic urban scenes with gaussian splatting. In European Conference on Computer Vision. 156–173

    2024

  36. [44]

    In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 2446–2454

  37. [45]

    Jizeng Wang, Xiaojing Liu, and Youhe Zhou. 2024. Application of wavelet methods in computational physics. Annalen der Physik 536, 5 (2024), 2300461

    2024

  38. [46]

    Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boning Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. 2024. Gps-gaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis. InProceedings of the IEEE/CVF Conference on Computer Vision and Patter...

    2024

  39. [47]

    Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming- Hsuan Yang. 2024. Drivinggaussian: Composite gaussian splatting for surround- ing dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognitio...

    2024

  40. [48]

    Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. 2023. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581 (2023)

    2023 arXiv

  41. [49]

    Jin-Chao Xu and Wei-Chang Shann. 1992. Galerkin-wavelet methods for two- point boundary value problems. Numer. Math. 63, 1 (1992), 123–144

    1992

  42. [51]

    Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. 2024. Mip-splatting: Alias-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 19447–19456

    2024

  43. [52]

    Zehao Yu, Torsten Sattler, and Andreas Geiger. 2024. Gaussian opacity fields: Efficient and compact surface reconstruction in unbounded scenes.arXiv preprint arXiv:2404.10772 (2024)

    2024 arXiv

  44. [56]

    Michael Zollhöfer, Patrick Stotko, Andreas Görlitz, et al. 2018. State of the art on 3D reconstruction with RGB-D cameras. In Computer graphics forum, Vol. 37. Wiley Online Library, 625–652. Received 20 February 2007; revised 12 March 2009; accepted 5 June 2009

    2018

  45. [2019]

    Advances in neural information processing systems 32 (2019)

    Advances in neural information processing systems. Advances in neural information processing systems 32 (2019)

    2019

  46. [2020]

    Advances in Neural Information Processing Systems 33 (2020), 15651–15663

    Neural sparse voxel fields. Advances in Neural Information Processing Systems 33 (2020), 15651–15663

    2020

  47. [2023]

    ACM Trans

    3D Gaussian splatting for real-time radiance field rendering. ACM Trans. Graph. 42, 4 (2023), 139–1

    2023

  48. [2024]

    In ACM SIGGRAPH 2024 conference papers

    2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 conference papers . 1–11

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.