Pith. sign in

REVIEW 4 major objections 6 minor 51 references

TraGraph-GS: Trajectory Graph-based Gaussian Splatting for Arbitrary Large-Scale Scene Rendering

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that partitioning large scenes by a camera-connectivity graph lets one 3D Gaussian Splatting pipeline render arbitrarily large aerial and ground scenes at state-of-the-art quality.

desk verdict The graph-partitioning idea is genuinely new and the aerial numbers are strong, but the ground-level SOTA claim is undercut by the paper's own table and missing baselines. read the letter →

arxiv 2506.08704 v1 pith:7RG5LJ3C submitted 2025-06-10 cs.CV

Xiaohan Zhang , Sitong Wang , Yushen Yan , Yi Yang , Mingda Xu , Qi Liu This is my paper
classification cs.CV
keywords 3DGaussianSplattinglarge-scalescenerenderingnovelviewsynthesistrajectorygraphgraph-basedspatialpartitioningmulti-viewconstraintmulti-scaleGaussiansprogressive
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

TraGraph-GS aims to show that a single divide-and-conquer pipeline can render arbitrarily large scenes, whether the cameras flew an orderly aerial grid or drove an irregular ground path, at state-of-the-art quality. Instead of cutting the scene by spatial grids or point-cloud density, it builds a graph in which images are nodes and the number of matched feature points between two images is the edge weight, then partitions nodes with a breadth-first search that keeps regions balanced and internally connected. Each region trains an independent 3D Gaussian Splatting model, regularized by a multi-view patch-consistency term and distance-aware multi-scale Gaussians, and novel views are rendered by progressively compositing the local models that predict the view best. The paper reports that this raises PSNR by 1.86 dB on average across four aerial datasets and 1.62 dB across four ground datasets relative to prior state-of-the-art methods, and that the same weights serve both scene types without per-dataset tuning.

What carries the argument

The load-bearing machinery is the camera connectivity graph $G=(V,E,w)$ built from Structure-from-Motion: images are vertices, and $w(u,v)$ is the feature-matching count, so high weights mean large co-visible areas. Around it are built (1) a BFS graph segmentation that grows each region from the highest-degree unassigned seed to $|V|/k$ nodes and then assigns leftover nodes to the least-loaded region whose score exceeds the median, (2) a multi-view constraint that warps patches between neighboring cameras by homography and penalizes low NCC, (3) a multi-scale densification rule that clones or splits Gaussians according to a distance factor $\gamma(\mu)=\max(1, \|\mu-c\|_2/r-1)$, and (4) a progressive renderer that accumulates projection opacity across local Gaussians sorted by matching score until the threshold $\beta$ is reached.

What would settle it

On a large-scale scene with a long textureless road or a repetitive facade, compute the SfM feature matches, build the graph, and compare the resulting subregions to a ground-truth co-visibility measure such as the fraction of shared match points between camera pairs; if clearly co-visible cameras are split into different regions while rendering quality drops, the central assumption fails.

Watch

Extended reading notes

Core claim

The central discovery the authors argue for is that the connectivity of the camera trajectory, read through feature-match counts, is the right spatial structure for partitioning large scenes. A breadth-first search over the image-connectivity graph yields subgraphs whose cameras share strong co-visibility, and this holds for dense uniform aerial grids, sparse irregular road drives, and mixed double-layer building captures alike; consequently the same code path replaces the specialized partitions of CityGaussian for aerial data and Hierarchy-GS for ground data. The paper further claims that two graph-enabled regularizers—a homography-based multi-view Normalized Cross Correlation constraint between neighboring cameras and a distance-scaled Gaussian densification for far objects—recover texture detail and distant geometry, and that a progressive rendering rule, which sorts local Gaussians by their matching score against a coarse global Gaussian and stops once projected opacity exceeds a threshold, removes the floating artifacts of naive merging. The evidence is the reported numbers: up to 3.42 dB PSNR over CityGaussian on Building, 4.82 dB over the ablated baseline on Road, and steadily improving quality with more regions.

Load-bearing premise

The whole method rests on the assumption that the number of matching feature points between two photos faithfully measures how much of the same scene they see, so the graph cut groups exactly the cameras that belong together.

Editorial extensions

If this is right

  • One pipeline now covers aerial, ground, and mixed trajectories without per-dataset tuning of the partitioning step, which previously required a separate method for each scene type.
  • Increasing the number of regions improves both quality and per-region training speed (the paper reports 47 minutes at 16 regions versus 1 hour 2 minutes at 2 regions), so the method scales by splitting further.
  • The progressive rendering module removes the floating-point artifacts caused by overlapping Gaussians at region boundaries, previously a standard failure of merged large-scene Gaussian models.
  • The accompanying depth and normal maps are cleaner than those of plain 3DGS, which the authors state is useful for scene measurement and mesh reconstruction.

Reading between the lines

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

  • The graph partition could be made more robust by learning co-visibility from appearance instead of raw feature-match counts; the paper does not explore this, but it would address the textureless-region failure mode.
  • Because the partition is trajectory-agnostic, the same pipeline should transfer to continuous video streams (for example, dashcam or drone footage) as long as SfM succeeds, though the paper only evaluates keyframe-style datasets.
  • The reported gains are averages; the largest margins come on the two lighting-stable aerial scenes, so the method's advantage likely narrows when appearance variation is strong, consistent with the paper's stated illumination limitation.
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

4 major / 6 minor

Summary. TraGraph-GS proposes a divide-and-conquer 3D Gaussian splatting pipeline for large-scale scenes with arbitrary camera trajectories. It builds a camera connectivity graph from SfM feature matching counts, segments the graph into k regions with a BFS-based algorithm, trains local 3DGS models per region with multi-view and multi-scale regularizers, trains a coarse global model, and renders new views by progressively accumulating local models sorted by a matching score. Experiments cover four aerial and four ground datasets, reporting PSNR/SSIM/LPIPS against NeRF and 3DGS baselines, plus ablations on the self-built Road dataset and a region-count study. The abstract claims average gains of 1.86 dB on aerial datasets and 1.62 dB on ground datasets compared to state-of-the-art approaches.

Significance. If the empirical claims were fully supported, the graph-based partitioning and progressive rendering components would be a practical contribution to large-scale scene rendering, particularly for trajectories that are not well handled by spatial grids or point-cloud partitions. The paper has useful strengths: it evaluates on four public aerial datasets, provides a component-wise ablation in Table III, and reports an efficiency/quality trade-off with region count in Table IV. However, the ground-level SOTA comparison is not supported by the reported numbers, two directly relevant prior methods are cited but not compared, and the progressive rendering equation lacks a validity argument. The core idea is worth pursuing, but the claims need to be corrected and the missing evidence supplied before the paper can be accepted.

major comments (4)
  1. [Abstract; Table II; Section IV-D] The abstract's claim of a 1.62 dB average improvement on ground datasets 'compared to state-of-the-art approaches' is not supported by Table II. The value 1.62 dB equals the mean of Ours minus CityGaussian across the four rows (3.87, 1.53, 0.38, and 0.68 dB), but Fig. 1 and Section II-B identify Hierarchy-GS, not CityGaussian, as the ground-level SOTA. Against Hierarchy-GS, TraGraph-GS loses on Campus (24.35 vs. 24.61 dB), and against the best baseline per dataset (Hierarchy-GS on SmallCity and Campus, CityGaussian on Road and General) the average advantage falls to roughly 0.43 dB. The two largest positive differences versus Hierarchy-GS occur on the self-built Road and General datasets, which are not publicly available. Please recompute the average against the best available baseline, state explicitly which baseline the quoted number refers to, or moderate the claim.
  2. [Section II-B/II-C and Tables I-II; Refs. [15,16]] TOY-GS [15] and Aerial-NeRF [16], both cited in the manuscript, address exactly the problem of partitioning and assembling local representations for large-scale free camera trajectories, yet neither is included in any quantitative or qualitative comparison. Because the title and abstract claim superiority over 'state-of-the-art approaches' for arbitrary trajectories, omitting these directly related methods leaves the central empirical claim incomplete. Please add them as baselines or provide a concrete justification for their exclusion.
  3. [Section III-C, Eqs. (10)-(11)] The progressive rendering rule in Eqs. (10)-(11) is not a valid alpha-compositing formula. The weight w_i is set to the projection opacity o_i(p) with no transmittance factor product_{j<i}(1-o_j) and no normalization, so the final color is a weighted sum of local renderings whose total weight can be below or above 1 (for example, o_1=0.8 with beta=0.5 leaves only 0.8 of any color with no background, while o_1=0.49, o_2=0.99 gives weights summing to 1.48). The manuscript provides no derivation that this produces correct colors, and the PR ablation in Table III does not isolate this blending rule from a standard front-to-back compositing alternative. Please replace Eq. (10) with a proper compositing expression or supply a correctness argument and an ablation comparing the two rules.
  4. [Section III-A, Algorithm 1] Algorithm 1 as written does not guarantee the balanced, connected regions that Section III-A claims. The BFS loop adds the entire adjacency set N at once ('P_i = P_i ∪ N'), so a region can overshoot the target size |V|/k before the while condition is rechecked; the final score-based assignment of unassigned nodes also enforces no size cap and can create regions of very different sizes. The pseudo-code needs a precise node-by-node BFS expansion with an explicit size constraint, or a correctness proof for the stated invariants. Since this segmentation is the component responsible for adapting to arbitrary camera trajectories, the discrepancy between the algorithm and its description should be resolved.
minor comments (6)
  1. [Section III-A and Algorithm 1] The word 'Camara' in the heading and algorithm caption should be 'Camera'.
  2. [Equations (9) and (11)] The summation index in Eq. (9) is overloaded with the region index i, and Eq. (11) writes the cumulative sum as 'i-1 P j=1 o_i(p)', where the subscript should be o_j(p).
  3. [Section IV-E] The text reports 'improvements of 4.82, 0.103, and 0.151 dB in PSNR, SSIM, and LPIPS', but only PSNR is measured in dB; SSIM and LPIPS are not logarithmic units. The same wording appears for Table IV and should be corrected.
  4. [Table I] The method name 'Swich-NeRF' should be 'Switch-NeRF' to match the text and the reference list.
  5. [Section IV-A/IV-D] The Road and General datasets are self-built and not released; please state their availability or provide a download link so the ground-level results can be reproduced, and clarify whether code will be released.
  6. [Section IV-B] The sentence 'we employ a single chunk for both training and test processes' is unclear, since test images are normally held out; please clarify what 'single chunk' means in the Hierarchy-GS protocol.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: TraGraph-GS is an empirical system evaluated on held-out views, and its self-citations are not load-bearing.

full rationale

TraGraph-GS is an empirical system, not a derivation. The spatial partitioning (Algorithm 1) consumes SfM feature-match counts and camera poses; the multi-view and multi-scale losses (Eqs. 5-7) are regularizers trained with photometric loss; and the progressive rendering (Eqs. 8-11) is a blending rule. None of these quantities is defined in terms of the reported PSNR/SSIM/LPIPS numbers, and the test metrics come from held-out images (every 8th image) on public benchmarks (Mill-19, UrbanScene3D, SmallCity, Campus) plus two self-built sets. The two self-citations ([15] TOY-GS and [16] Aerial-NeRF) appear only in the related-work sentence 'These methods [1] [14] [2] [15] [16] supported parallel processing of individual chunks followed by consolidation'; they are not used to justify the method or to forbid alternatives, so they are not load-bearing. The abstract's 1.62 dB ground average is arithmetically the mean gain over CityGaussian even though the paper calls Hierarchy-GS the ground-level SOTA; that is a baseline-consistency or correctness concern, not a circular reduction of a prediction to a fit. No equation or fitted parameter is renamed as a prediction. Hence no significant circularity.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The method relies on standard 3DGS machinery plus hand-set hyperparameters. The nonstandard assumptions are the graph co-visibility proxy and the validity of the progressive blending formula. No new physical or conceptual entities are introduced; the trajectory graph, local Gaussians, and coarse global Gaussian are all standard representations.

free parameters (5)
  • number of regions k = 8 (aerial), 3 (self-captured), 1 (SmallCity, Campus)
    Set per dataset type rather than derived; Table IV shows PSNR varies from 25.81 to 28.96 as k goes from 2 to 16, so the choice materially affects results.
  • loss weight lambda = 0.2
    Fixed weight balancing L1/SSIM and multi-view loss in Eq. (6) and Eq. (8); no sensitivity analysis is provided.
  • progressive rendering threshold beta = 0.5
    Threshold in Eq. (11) controlling how many local Gaussians contribute per pixel; chosen by hand without a search or ablation.
  • multi-scale thresholds g and b = g=0.01, b=0.1
    Controls clone, split, and removal of Gaussians at distance in Eq. (7); values are fixed manually.
  • number of neighbor cameras for multi-view constraint = 4
    The four highest-edge-weight neighbors are used for the NCC loss; no ablation on this number is provided.
assumptions (6)
  • domain assumption SfM provides accurate camera poses and sparse point clouds for the input images.
    Used in Sec. III-A to build the graph and initialize Gaussians; if poses are inaccurate, partitioning and rendering fail.
  • domain assumption Feature matching counts between image pairs are a faithful measure of co-visibility.
    Central to the graph in Sec. III-A; the whole partitioning depends on this proxy.
  • standard math The BFS graph segmentation algorithm produces balanced, connected subgraphs and terminates as intended.
    Algorithm 1 assumes standard BFS behavior, but the pseudocode as written may not terminate cleanly because nodes are added until the region reaches |V|/k without recomputing N from the updated region.
  • domain assumption The patch-matching NCC loss in Eq. (5) improves geometry rather than just overfitting appearance.
    The multi-view constraint relies on this prior; in textureless regions the plane-sweep homography may not be reliable.
  • ad hoc to paper The progressive rendering weighting in Eqs. (10)-(11) yields a valid color blend.
    The weights are not normalized, and no derivation shows that summing opacity-weighted colors until a threshold equals a proper alpha-composited result.
  • domain assumption Position-aware adaptive control from Periodic Vibration Gaussian improves distant rendering when applied inside each local region.
    The multi-scale constraint is taken from [50] and assumed to transfer to the partitioned setting.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TraGraph-GS: Trajectory Graph-based Gaussian Splatting for Arbitrary Large-Scale Scene Rendering." pith.science (2026). https://pith.science/paper/7RG5LJ3C

@misc{pith2026250608704,
  author       = {Pith},
  title        = {Pith review of: TraGraph-GS: Trajectory Graph-based Gaussian Splatting for Arbitrary Large-Scale Scene Rendering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7RG5LJ3C}},
  note         = {Machine review of arXiv:2506.08704}
}
read the original abstract

High-quality novel view synthesis for large-scale scenes presents a challenging dilemma in 3D computer vision. Existing methods typically partition large scenes into multiple regions, reconstruct a 3D representation using Gaussian splatting for each region, and eventually merge them for novel view rendering. They can accurately render specific scenes, yet they do not generalize effectively for two reasons: (1) rigid spatial partition techniques struggle with arbitrary camera trajectories, and (2) the merging of regions results in Gaussian overlap to distort texture details. To address these challenges, we propose TraGraph-GS, leveraging a trajectory graph to enable high-precision rendering for arbitrarily large-scale scenes. We present a spatial partitioning method for large-scale scenes based on graphs, which incorporates a regularization constraint to enhance the rendering of textures and distant objects, as well as a progressive rendering strategy to mitigate artifacts caused by Gaussian overlap. Experimental results demonstrate its superior performance both on four aerial and four ground datasets and highlight its remarkable efficiency: our method achieves an average improvement of 1.86 dB in PSNR on aerial datasets and 1.62 dB on ground datasets compared to state-of-the-art approaches.

Figures

Figures reproduced from arXiv: 2506.08704 by the authors.

Figure 1
Figure 1. Existing methods have shown superior performance in specific sce [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Pipeline of our TraGraph-GS: a) Scene Partitioning: From unordered images, we first extract features and construct a camera connectivity graph [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Trajectories of our created datasets. The upper figure shows the long [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Visualization results on four aerial datasets. Aerial scenes present significant challenges in accurately rendering texture details of ground-level objects. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Comparison of scene rendering and structural information. Our approach offers enhanced scene rendering and more accurate structural information [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Visualization results of four ground datasets. Unbounded scene rendering from a ground-level perspective is inherently challenging. In contrast to [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Visualization of spatial partitioning based on our graph segmentation. (a) Partitioning cameras and point clouds of large-scale aerial scenes captured [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Visualization results of our ablation experiments. Comparing (a) [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 25 canonical work pages

  1. [15]

    Toy-gs: Assembling local gaussians for precisely rendering large-scale free camera trajectories,

    X. Zhang, Z. Sun, Y . Qiu, J. Su, and Q. Liu, “Toy-gs: Assembling local gaussians for precisely rendering large-scale free camera trajectories,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 10, 2025, pp. 10 212–10 220

    work page 2025

  2. [16]

    Aerial-nerf: Adaptive spatial par- titioning and sampling for large-scale aerial rendering,

    X. Zhang, Y . Qiu, Z. Sun, and Q. Liu, “Aerial-nerf: Adaptive spatial par- titioning and sampling for large-scale aerial rendering,”arXiv preprint arXiv:2405.06214, 2024

    arXiv 2024

  3. [1]

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

    Y . Liu, C. Luo, L. Fan, N. Wang, J. Peng, and Z. Zhang, “Citygaussian: Real-time high-quality large-scale scene rendering with gaussians,” in European Conference on Computer Vision. Springer, 2024, pp. 265– 282

    work page 2024

  4. [2]

    A hierarchical 3d gaussian representation for real-time rendering of very large datasets,

    B. Kerbl, A. Meuleman, G. Kopanas, M. Wimmer, A. Lanvin, and G. Drettakis, “A hierarchical 3d gaussian representation for real-time rendering of very large datasets,”ACM Transactions on Graphics (TOG), vol. 43, no. 4, pp. 1–15, 2024

    2024

  5. [3]

    Modeling and rendering architecture from photographs: A hybrid geometry-and image-based ap- proach,

    P. E. Debevec, C. J. Taylor, and J. Malik, “Modeling and rendering architecture from photographs: A hybrid geometry-and image-based ap- proach,” inSeminal Graphics Papers: Pushing the Boundaries, Volume 2, 2023, pp. 465–474

    work page 2023

  6. [4]

    Un- structured lumigraph rendering,

    C. Buehler, M. Bosse, L. McMillan, S. Gortler, and M. Cohen, “Un- structured lumigraph rendering,” inProceedings of the 28th annual conference on Computer graphics and interactive techniques, 2001, pp. 425–432

    work page 2001

  7. [5]

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

    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,”Communications of the ACM, vol. 65, no. 1, pp. 99–106, 2021

    2021

  8. [6]

    Mip-nerf: A multiscale representation for anti- aliasing neural radiance fields,

    J. T. Barron, B. Mildenhall, M. Tancik, P. Hedman, R. Martin-Brualla, and P. P. Srinivasan, “Mip-nerf: A multiscale representation for anti- aliasing neural radiance fields,” inProceedings of the IEEE/CVF inter- national conference on computer vision, 2021, pp. 5855–5864

    2021

Show all 51 references
  1. [7]

    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,” inPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 5470–5479

    2022

  2. [8]

    Instant neural graphics primitives with a multiresolution hash encoding,

    T. M ¨uller, A. Evans, C. Schied, and A. Keller, “Instant neural graphics primitives with a multiresolution hash encoding,”ACM transactions on graphics (TOG), vol. 41, no. 4, pp. 1–15, 2022

    2022

  3. [9]

    Tensorf: Tensorial radiance fields,

    A. Chen, Z. Xu, A. Geiger, J. Yu, and H. Su, “Tensorf: Tensorial radiance fields,” inEuropean conference on computer vision. Springer, 2022, pp. 333–350

    2022

  4. [10]

    Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps,

    C. Reiser, S. Peng, Y . Liao, and A. Geiger, “Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 14 335–14 345

    2021

  5. [11]

    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

    2023

  6. [12]

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

    Z. Yu, A. Chen, B. Huang, T. Sattler, and A. Geiger, “Mip-splatting: Alias-free 3d gaussian splatting,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 19 447–19 456

    2024

  7. [13]

    Scaffold-gs: Structured 3d gaussians for view-adaptive rendering,

    T. Lu, M. Yu, L. Xu, Y . Xiangli, L. Wang, D. Lin, and B. Dai, “Scaffold-gs: Structured 3d gaussians for view-adaptive rendering,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 20 654–20 664

    2024

  8. [14]

    Vastgaussian: Vast 3d gaussians for large scene reconstruction,

    J. Lin, Z. Li, X. Tang, J. Liu, S. Liu, J. Liu, Y . Lu, X. Wu, S. Xu, Y . Yan et al., “Vastgaussian: Vast 3d gaussians for large scene reconstruction,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 5166–5175

    2024

  9. [17]

    View morphing,

    S. M. Seitz and C. R. Dyer, “View morphing,” inProceedings of the 23rd annual conference on Computer graphics and interactive techniques, 1996, pp. 21–30

    1996

  10. [18]

    Nerf++: Analyzing and improving neural radiance fields,

    K. Zhang, G. Riegler, N. Snavely, and V . Koltun, “Nerf++: Analyzing and improving neural radiance fields,”arXiv preprint arXiv:2010.07492, 2020

    2010 arXiv

  11. [19]

    Nerf in the wild: Neural radiance fields for unconstrained photo collections,

    R. Martin-Brualla, N. Radwan, M. S. Sajjadi, J. T. Barron, A. Doso- vitskiy, and D. Duckworth, “Nerf in the wild: Neural radiance fields for unconstrained photo collections,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 7210– 7219

    2021

  12. [20]

    Plenoxels: Radiance fields without neural networks,

    S. Fridovich-Keil, A. Yu, M. Tancik, Q. Chen, B. Recht, and A. Kanazawa, “Plenoxels: Radiance fields without neural networks,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 5501–5510

    2022

  13. [21]

    Neural sparse voxel fields,

    L. Liu, J. Gu, K. Zaw Lin, T.-S. Chua, and C. Theobalt, “Neural sparse voxel fields,”Advances in Neural Information Processing Systems, vol. 33, pp. 15 651–15 663, 2020

    2020

  14. [22]

    K-planes: Explicit radiance fields in space, time, and appearance,

    S. Fridovich-Keil, G. Meanti, F. R. Warburg, B. Recht, and A. Kanazawa, “K-planes: Explicit radiance fields in space, time, and appearance,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 12 479–12 488

    2023

  15. [23]

    Tri- miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields,

    W. Hu, Y . Wang, L. Ma, B. Yang, L. Gao, X. Liu, and Y . Ma, “Tri- miprf: Tri-mip representation for efficient anti-aliasing neural radiance fields,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 19 774–19 783

    2023

  16. [24]

    Giganticnvs: Gigapixel large-scale neural rendering with implicit meta-deformed manifold,

    G. Wang, J. Zhang, K. Zhang, R. Huang, and L. Fang, “Giganticnvs: Gigapixel large-scale neural rendering with implicit meta-deformed manifold,”IEEE Transactions on Pattern Analysis and Machine Intelli- gence, pp. 338–353, 2024

    2024

  17. [25]

    Editablenerf: Editing topologically varying neural radiance fields by key points,

    C. Zheng, W. Lin, and F. Xu, “Editablenerf: Editing topologically varying neural radiance fields by key points,”IEEE Transactions on Pattern Analysis and Machine Intelligence, pp. 5779–5790, 2024

    2024

  18. [26]

    Ltm-nerf: Embedding 3d local tone mapping in hdr neural radiance field,

    X. Huang, Q. Zhang, Y . Feng, H. Li, and Q. Wang, “Ltm-nerf: Embedding 3d local tone mapping in hdr neural radiance field,”IEEE Transactions on Pattern Analysis and Machine Intelligence, pp. 10 944– 10 959, 2024

    2024

  19. [27]

    Vd-nerf: Visibility-aware decoupled neural radiance fields for view-consistent editing and high- frequency relighting,

    T. Wu, J.-M. Sun, Y .-K. Lai, and L. Gao, “Vd-nerf: Visibility-aware decoupled neural radiance fields for view-consistent editing and high- frequency relighting,”IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, pp. 3344–3357, 2025

    2025

  20. [28]

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

    Z. Fan, K. Wang, K. Wen, Z. Zhu, D. Xu, Z. Wanget al., “Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps,”Advances in neural information processing systems, vol. 37, pp. 140 138–140 158, 2024. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 12

    2024

  21. [29]

    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

    2024

  22. [30]

    Deferredgs: Decoupled and relightable gaussian splatting with deferred shading,

    T. Wu, J.-M. Sun, Y .-K. Lai, Y . Ma, L. Kobbelt, and L. Gao, “Deferredgs: Decoupled and relightable gaussian splatting with deferred shading,” IEEE Transactions on Pattern Analysis and Machine Intelligence, pp. 1–13, 2025

    2025

  23. [31]

    Block-nerf: Scalable large scene neural view synthesis,

    M. Tancik, V . Casser, X. Yan, S. Pradhan, B. Mildenhall, P. P. Srinivasan, J. T. Barron, and H. Kretzschmar, “Block-nerf: Scalable large scene neural view synthesis,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 8248–8258

    2022

  24. [32]

    Gaussianpro: 3d gaussian splatting with progressive propa- gation,

    K. Cheng, X. Long, K. Yang, Y . Yao, W. Yin, Y . Ma, W. Wang, and X. Chen, “Gaussianpro: 3d gaussian splatting with progressive propa- gation,” inForty-first International Conference on Machine Learning, 2024

    2024

  25. [33]

    Luebke, M

    D. Luebke, M. Reddy, J. D. Cohen, A. Varshney, B. Watson, and R. Huebner,Level of detail for 3D graphics. Elsevier, 2002

    2002

  26. [34]

    Sequential point trees,

    C. Dachsbacher, C. V ogelgsang, and M. Stamminger, “Sequential point trees,”ACM Transactions on Graphics (TOG), vol. 22, no. 3, pp. 657– 662, 2003

    2003

  27. [35]

    Qsplat: A multiresolution point ren- dering system for large meshes,

    S. Rusinkiewicz and M. Levoy, “Qsplat: A multiresolution point ren- dering system for large meshes,” inProceedings of the 27th annual conference on Computer graphics and interactive techniques, 2000, pp. 343–352

    2000

  28. [36]

    Multi-scale 3d gaussian splatting for anti-aliased rendering,

    Z. Yan, W. F. Low, Y . Chen, and G. H. Lee, “Multi-scale 3d gaussian splatting for anti-aliased rendering,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 20 923–20 931

    2024

  29. [37]

    Photo tourism: exploring photo collections in 3d,

    N. Snavely, S. M. Seitz, and R. Szeliski, “Photo tourism: exploring photo collections in 3d,” inACM siggraph 2006 papers, 2006, pp. 835–846

    2006

  30. [38]

    Building rome in a day,

    S. Agarwal, Y . Furukawa, N. Snavely, I. Simon, B. Curless, S. M. Seitz, and R. Szeliski, “Building rome in a day,”Communications of the ACM, vol. 54, no. 10, pp. 105–112, 2011

    2011

  31. [39]

    Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs,

    H. Turki, D. Ramanan, and M. Satyanarayanan, “Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 12 922–12 931

    2022

  32. [40]

    Switch-nerf: Learning scene decomposition with mixture of experts for large-scale neural radiance fields,

    M. Zhenxing and D. Xu, “Switch-nerf: Learning scene decomposition with mixture of experts for large-scale neural radiance fields,” inThe Eleventh International Conference on Learning Representations, 2022

    2022

  33. [41]

    Urban radiance fields,

    K. Rematas, A. Liu, P. P. Srinivasan, J. T. Barron, A. Tagliasacchi, T. Funkhouser, and V . Ferrari, “Urban radiance fields,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 12 932–12 942

    2022

  34. [42]

    Dogs: Distributed-oriented gaussian splatting for large-scale 3d reconstruction via gaussian consensus,

    Y . Chen and G. H. Lee, “Dogs: Distributed-oriented gaussian splatting for large-scale 3d reconstruction via gaussian consensus,”Advances in Neural Information Processing Systems, vol. 37, pp. 34 487–34 512, 2024

    2024

  35. [43]

    Distinctive image features from scale-invariant keypoints,

    D. G. Lowe, “Distinctive image features from scale-invariant keypoints,” International journal of computer vision, vol. 60, pp. 91–110, 2004

    2004

  36. [44]

    Structure-from-motion revisited,

    J. L. Schonberger and J.-M. Frahm, “Structure-from-motion revisited,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 4104–4113

    2016

  37. [45]

    Algebraic connectivity of graphs,

    M. Fiedler, “Algebraic connectivity of graphs,”Czechoslovak mathemat- ical journal, vol. 23, no. 2, pp. 298–305, 1973

    1973

  38. [46]

    An efficient heuristic procedure for partitioning graphs,

    B. W. Kernighan and S. Lin, “An efficient heuristic procedure for partitioning graphs,”The Bell system technical journal, vol. 49, no. 2, pp. 291–307, 1970

    1970

  39. [47]

    Breadth-first search,

    A. Bundy and L. Wallen, “Breadth-first search,”Catalogue of artificial intelligence tools, pp. 13–13, 1984

    1984

  40. [48]

    Patch- match: A randomized correspondence algorithm for structural image editing,

    C. Barnes, E. Shechtman, A. Finkelstein, and D. B. Goldman, “Patch- match: A randomized correspondence algorithm for structural image editing,”ACM Trans. Graph., vol. 28, no. 3, p. 24, 2009

    2009

  41. [49]

    Fast normalized cross-correlation,

    J.-C. Yoo and T. H. Han, “Fast normalized cross-correlation,”Circuits, systems and signal processing, vol. 28, pp. 819–843, 2009

    2009

  42. [50]

    Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering,

    Y . Chen, C. Gu, J. Jiang, X. Zhu, and L. Zhang, “Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering,” arXiv preprint arXiv:2311.18561, 2023

    2023 arXiv

  43. [51]

    Capturing, reconstructing, and simulating: the urbanscene3d dataset,

    L. Lin, Y . Liu, Y . Hu, X. Yan, K. Xie, and H. Huang, “Capturing, reconstructing, and simulating: the urbanscene3d dataset,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 93–109. Xiaohan Zhangreceived the M.Sc. degree from the School of Mathematics and Statist...

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.