Pith. sign in

REVIEW 2 major objections 6 minor 49 references

The paper claims that a coarse, first-order SfM initialization suffices for high-quality 3D Gaussian Splatting when combined with joint pose-and-appearance refinement and an importance-guided MCMC Gaussian allocation that steers a fixed bud

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 06:56 UTC pith:ZJU2VRIM

load-bearing objection A genuinely transparent 3DGS systems paper with a modest but real allocation heuristic; the headline macro-average lead is softer than it looks and needs disclosure fixes before it is acceptably comparable. the 2 major comments →

arxiv 2607.11285 v2 pith:ZJU2VRIM submitted 2026-07-13 cs.CV

SalientGS: Unified SfM-to-3DGS with Importance-Guided MCMC Gaussian Allocation

Tianyu Xiong , Rui Li , Suning Ge , Jiaqi Yang This is my paper
classification cs.CV
keywords 3D Gaussian SplattingStructure from MotionMCMCGaussian AllocationNovel View SynthesisJoint Pose OptimizationBundle AdjustmentImportance Sampling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

SalientGS sets out to prove that 3D Gaussian Splatting does not need the expensive, second-order Structure-from-Motion (SfM) preprocessing that most current pipelines depend on. The paper builds a unified pipeline in which a fast first-order SfM stage, cheap descriptor-based image matching, and joint refinement of camera poses and the Gaussian representation replace the traditional frozen SfM front end. Its central mechanism is an importance-guided Markov Chain Monte Carlo allocation step: per-Gaussian underfit and redundancy scores are aggregated from multi-view rendering errors, and a smooth sampling distribution biases Gaussian birth and relocation toward underfit regions under a fixed 1.5-million budget. The paper reports the best cross-benchmark macro-average PSNR/SSIM/LPIPS among compared methods, the fastest end-to-end time (about 10-12 minutes across benchmarks), and pose-accuracy results on 45 real sequences showing that a weak first-order SfM initialization plus joint optimization surpasses stronger SfM baselines. The broader claim is that attentive capacity reallocation makes a coarse initialization sufficient, shifting the bottleneck from SfM to the allocation and refinement machinery.

Core claim

The paper claims that a coarse, first-order SfM initialization is sufficient for high-quality 3D Gaussian Splatting when camera poses are jointly refined with the representation and a fixed Gaussian budget is steered by multi-view residuals. Its key mechanism, importance-guided MCMC allocation, aggregates per-pixel rendering errors into per-Gaussian underfit and redundancy scores, then uses a smooth importance-weighted distribution to bias Gaussian birth and relocation toward underfit regions. On 13 scenes across three benchmarks, the method reports the best macro-average PSNR/SSIM/LPIPS and the fastest end-to-end time; on 45 pose-accuracy sequences, a weak initialization plus joint optimiza

What carries the argument

The load-bearing object is an importance-weighted sampling distribution built from per-Gaussian scores. During training, several views are rendered and per-pixel L1 error maps are computed; robust quantile normalization turns these into underfit masks. Each Gaussian's importance score is the average normalized overlap of its 2D footprint with high-error masks across views, and a redundancy score is the analogous overlap with low-error masks. A softplus-transformed weight, mixed with opacity to encourage exploration, forms the proposal distribution used both to teleport low-value Gaussians and to spawn new ones. This sits on top of an MCMC population-management framework, leaving its underlyi

Load-bearing premise

The method relies on the assumption that errors from a cheap first-order SfM stage are smooth and systematic rather than discrete outliers, so that photometric and reprojection gradients can correct them during joint training; if the initial poses contain gross outliers or the view graph is poorly connected, the geometric anchor can lock training onto wrong geometry and the claimed recovery fails.

What would settle it

Run the pipeline on a scene where the first-order SfM produces a clearly disconnected or degenerate view graph (e.g., a long corridor with collinear camera motion or severe appearance change between overlapping views) and check whether joint pose optimization can recover; the paper itself reports that two of 45 ETH3D sequences fail SfM entirely and that reducing image retrieval to k=5 makes the stump scene collapse to 14.97 dB PSNR, so a decisive test would be to identify the scale of pose error beyond which the BA anchor cannot recover.

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

If this is right

  • End-to-end reconstruction time on unordered image collections can drop by roughly an order of magnitude without sacrificing rendering quality, because the costly global SfM stage is replaced by a fast first-order one.
  • Fixed-budget methods become more attractive: the same number of Gaussians achieves higher perceptual quality when capacity is steered by multi-view error, so smaller models can match larger ones.
  • Joint pose-and-appearance optimization with an explicit reprojection anchor is a viable way to correct systematic errors in cheap SfM, suggesting that expensive refinement can be moved into the training loop.
  • Robustness across benchmarks improves: the method claims top-two LPIPS on every benchmark, whereas some pose-optimizing competitors fail catastrophically on certain scenes.

Where Pith is reading between the lines

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

  • The mechanism is not tied to Gaussian Splatting specifically; the idea of attributing multi-view residual error to individual primitives and using it to bias allocation could be applied to other capacity-limited scene representations, such as point-based or voxel-based radiance fields.
  • The paper's finding that a weak first-order initialization plus joint optimization beats stronger initializations suggests that the real value of SfM may be a reliable view graph rather than highly accurate poses, which would reframe where research should focus.
  • A testable extension: use learned descriptors in the retrieval stage to test whether robustness to appearance variation improves while keeping the same allocation machinery; the paper notes its descriptor-based retrieval may fail under severe appearance changes.
  • The importance guidance gains are largest at small budgets (0.27 dB at 500K vs 0.09 dB at 3M), implying the method is especially suited to memory-constrained deployment; one could push this further to very tight budgets.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 6 minor

Summary. The paper proposes SalientGS, an end-to-end SfM-to-3D Gaussian Splatting pipeline that replaces the usual COLMAP preprocessing with Fisher Vector + MST matching, first-order SfM, and joint photometric plus bundle-adjustment pose optimization during 3DGS-MCMC training. The central algorithmic contribution is importance-guided MCMC Gaussian allocation: per-Gaussian underfit and redundancy scores are derived from multi-view L1 error maps and are used to bias birth and relocation proposals toward underfit regions, without modifying the underlying SGLD updates. On 13 benchmark scenes (Mip-NeRF 360, Deep Blending, Tanks & Temples) with released code and a fixed 30K-step schedule, the method reports the best three-benchmark macro-average PSNR/SSIM/LPIPS (27.65 dB / 0.876 / 0.147) and the fastest end-to-end runtime (10.62 min) among the compared methods. Ablations quantify the contributions of guided MCMC, joint pose optimization, and SfM initialization quality, and an ETH3D pose study compares FastMap, COLMAP, and GLOMAP before and after joint optimization.

Significance. If the main claims hold, the paper demonstrates that a fast, coarse first-order SfM front-end can be made sufficient for high-quality fixed-budget 3DGS when paired with joint photometric/BA pose refinement, and that MCMC allocation can be steered by a multi-view error signal without changing the SGLD population dynamics. The paper has notable strengths: released code and evaluation scripts, per-scene measurements, explicit disclosure of a prior evaluation error, dedicated failure-case analysis, and an unusually honest limitations section. The individual contribution of the proposed allocation is modest (+0.10 dB PSNR over vanilla MCMC at 1.5M Gaussians, Table 2), but the integrated pipeline is practically interesting and the runtime argument is compelling. However, the evidence for the central 'coarse SfM is sufficient' premise is incomplete: the error model for first-order SfM is asserted rather than demonstrated, and the benchmark lacks a same-pipeline SfM-backend control. These issues are fixable with additional experiments and analysis.

major comments (2)
  1. [§4.6 / Eq. (11) / Table 5] The 'smooth and systematic' error model for first-order SfM is stated in the text but not demonstrated by the evidence shown. Table 5 aggregates pose errors over 39 or 34 successfully processed sequences; aggregate Recall/AUC values cannot distinguish a handful of large outliers from uniformly small errors. The paper's own numbers show the failure regime is not hypothetical (2/45 SfM failures, 4/45 joint-training failures, and k=5 retrieval degrading stump to 14.97 dB in Table 3). Since Eq. (11) fixes the triangulated track points X_k at SfM initialization and optimizes only camera poses, a systematic bias in the initial tracks could be locked in or even reinforced by the BA anchor, and photometric refinement would then have to fight that anchor. I request: (i) per-sequence/per-camera pose error distributions before and after joint optimization for FastMap, COLMAP, and GLOMAP on ETH3D; (
  2. [§4.2 / Table 1] The headline macro-average includes failed drjohnson reconstructions for GloSplat-A and VGGT-X. Inclusion is a legitimate and transparent protocol choice, and the paper explicitly states that failed scenes are not dropped. Nevertheless, the claim that SalientGS has the best three-benchmark macro-average would be considerably strengthened by a sensitivity analysis: report scene-median metrics and macro-averages computed after (a) excluding methods' failed scenes and (b) replacing the failed scene with the next-worst non-failed scene. The released per-scene CSV presumably makes this possible, and it would show whether the macro-average lead survives a less catastrophic comparison. In addition, all verification runs use a single seed (seed 42) on a single GPU; for a stochastic MCMC-based allocation method, one or two additional seeds would materially strengthen the reproducibility claim.
minor comments (6)
  1. [Abstract / §4.2] The abstract says 'reconstruction in 15 minutes', while Table 1 reports a 10.62-minute average. Please reconcile these numbers.
  2. [Eq. (9)] The normalization of \tilde{w}_i to unit mean before multiplying by the opacity-mixing term is underspecified. State the exact order of operations, and whether normalization happens before or after applying the (1-\lambda_mix)\sigma(o_i)+\lambda_mix factor.
  3. [Figure 2] The diagram contains garbled text 'VeNNcaNon'; this should presumably read 'Verification'. Please fix all OCR-like artifacts in figures.
  4. [Table 5] Clarify how the 39 'Train-succ' and 34 'Recon succ' subsets are defined and whether the same subsets are used for all three SfM methods. The text mentions 2 SfM failures and 4 joint-training failures; explicitly state whether these are disjoint and how they map to 43, 39, and 34.
  5. [§4.3] The sentence 'A previous evaluation error affected only the reported full-model aggregate' is intriguing but unexplained. Provide enough detail or a clear pointer to the supplementary material so readers understand the nature of the error and the correction.
  6. [Related Work / Table 1] GloSplat [41] is the authors' prior work and is used as a key pose-optimizing baseline. This relationship should be stated explicitly in the main text near the comparison, rather than only being inferable from the author list of the reference.

Circularity Check

0 steps flagged

No significant circularity: importance-guided allocation is a heuristic layered on 3DGS-MCMC, and the pose-refinement claims rest on external benchmarks and ablations rather than on a self-referential derivation.

full rationale

SalientGS's central contribution is an importance-weighted sampling distribution computed from multi-view L1 error maps (Eqs. 2–9). The paper explicitly labels this a 'heuristic allocation strategy' that leaves the underlying SGLD parameter updates unchanged and 'makes no additional convergence claims' (Section 3.3), so no derived quantity is equivalent to its input by construction. The pose-refinement claim is supported by the ETH3D cross-product experiment (Table 5) comparing FastMap, COLMAP, and GLOMAP before and after joint optimization; this is an externally falsifiable empirical comparison, not a fitted parameter renamed as a prediction. The only self-citation is GloSplat [41], used as a baseline; its drjohnson failure is reported transparently and is not used to justify the central derivation. The 'smooth and systematic' error assumption (Section 4.6) is an empirical hypothesis that may be under-supported given aggregate metrics and the reported failure cases, but that is a correctness/robustness concern rather than circularity. The paper's own limitation statement — 'our joint optimization refines rather than substitutes for successful SfM initialization' (Section 5) — further confirms that the pipeline does not assume its conclusion. No equation reduces to another by construction, and no load-bearing uniqueness theorem is imported from the authors' prior work.

Axiom & Free-Parameter Ledger

12 free parameters · 7 axioms · 0 invented entities

The central claim is an empirical systems claim, so there are no mathematical axioms in the proof sense. The entries above are the unproven domain assumptions and hand-set design choices that the headline numbers depend on.

free parameters (12)
  • importance threshold τimp = 5 (0–100 scale)
    Hand-set threshold in Eq. 9 defining how much persistent underfit mass is needed to prioritize a Gaussian; central to importance-weighting.
  • redundancy threshold τred = 0.9
    Hand-set in Section 4.1; Gaussians with min-max-normalized redundancy score > 0.9 are relocation candidates.
  • opacity mixing weight λmix = 0.05
    Eq. 9; prevents under-sampling of low-opacity but underfit Gaussians; chosen by hand.
  • view sample count K = 10
    Number of views used per score recomputation; Section 4.1 and Eq. 6.
  • normalization quantiles (ℓ,h) = (0.05, 0.90)
    Robust quantile normalization in Eq. 3; one setting across all scenes.
  • selection quantiles (q_hi,q_lo) = (0.9, 0.1)
    Per-view thresholds in Eq. 4.
  • top-k retrieval = 20
    Number of candidate image pairs per query; ablation shows k=5 costs 2.25 dB and k=40 gives +0.04 dB (Table 3).
  • GMM components M = 64
    Fisher Vector GMM size; Section 4.1.
  • Gaussian budget cap N_GS = 1.5M
    Fixed budget is central to the claim; guidance gains vary with budget (+0.27 dB at 500K to +0.09 dB at 3M, Table 4).
  • BA/photometric loss weights λ_BA, λ_s = 0.01, 0.2
    Loss weights in Eq. 10–12; chosen by hand.
  • score recompute period T and warmup = T=500, warmup=3K
    Recompute importance scores every 500 iterations after 3K pose warmup; Section 4.1.
  • training iterations = 30K
    Exact 30K step schedule; fixed for all scenes.
axioms (7)
  • domain assumption Base 3DGS-MCMC SGLD population dynamics are correct and unchanged; importance guidance adds no convergence guarantee
    Section 3.3 states guidance is a heuristic allocation layer atop SGLD; if the base MCMC framework is flawed, SalientGS inherits the flaw.
  • domain assumption Per-pixel L1 photometric error is a reliable proxy for where additional Gaussian capacity improves perceptual quality
    Importance/redundancy scores are built from L1 maps (Eqs. 2–8); there is no direct evidence this aligns with LPIPS, and Section 4.5 shows a per-view LPIPS failure on room despite the pipeline's overall performance.
  • domain assumption First-order SfM errors are smooth and systematic, not discrete outliers, so joint photometric+BA refinement can correct them
    Stated in Section 4.6 as the explanation for why FastMap+joint optimization surpasses GLOMAP+joint; the whole fast-SfM pipeline rests on this.
  • domain assumption SfM triangulated tracks and their 2D associations remain reliable enough to anchor BA during joint training
    Eq. 11 freezes track points at SfM initialization; if tracks are contaminated, BA could anchor poses to wrong geometry.
  • domain assumption Fisher Vector retrieval + MST produces a connected, reliable view graph with top-k=20
    Section 3.1; ablation shows k=5 causes catastrophic failure on stump (Table 3), so the guarantee is load-bearing.
  • domain assumption Macro-averaging three dataset-level metrics and retaining failed baseline scenes is a fair comparison protocol
    Section 4.1; the headline PSNR/SSIM/LPIPS and runtime advantages depend on this aggregation choice.
  • standard math Fixing camera 0 resolves gauge ambiguity
    Section 3.4; standard fix for the gauge freedom in joint pose optimization.

pith-pipeline@v1.3.0-alltime-deepseek · 13636 in / 17820 out tokens · 147750 ms · 2026-08-02T06:56:44.479715+00:00 · methodology

0 comments
read the original abstract

Reconstructing 3D scenes from unordered images remains bottlenecked by expensive Structure-from-Motion (SfM) preprocessing and frozen pose interfaces. We present SalientGS, a unified SfM-to-3D Gaussian Splatting (3DGS) pipeline. Its central contribution is importance-guided Markov Chain Monte Carlo (MCMC) Gaussian allocation, which aggregates multi-view residuals into per-Gaussian underfit and redundancy signals. These signals define a smooth importance-weighted sampling distribution that biases both birth and relocation toward underfit regions. This reallocates capacity from well-fit areas without altering the underlying stochastic gradient Langevin dynamics (SGLD). SalientGS achieves end-to-end reconstruction in 15 minutes with state-of-the-art perceptual quality. The supplementary material provides dedicated sections for Per-Scene Qualitative Comparisons and Per-Image Learned Perceptual Image Patch Similarity (LPIPS) Analysis, including failure cases. Code and evaluation scripts are available at https://github.com/Six-Bit-TX/SalientGS.

Figures

Figures reproduced from arXiv: 2607.11285 by Jiaqi Yang, Rui Li, Suning Ge, Tianyu Xiong.

Figure 1
Figure 1. Figure 1: Quality vs. Speed vs. Model Size on Mip-NeRF 360. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Starting from an unordered image set, we first retrieve candidate pairs with Fisher Vector descriptors and enforce [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Given an unordered image collection, we extract [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: SfM runtime vs. image count. SalientGS achieves [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Selected-view qualitative comparison on Mip-NeRF 360. The first three panels illustrate localized detail differences on [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 3 linked inside Pith

  1. [1]

    Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic

  2. [2]

    Gabriele Barbarani, Gabriele Trivigno, Carlo Berton, Debora Mereu, and Carlo Masone. 2024. MegaLoc: One Retrieval to Place Them All. InEuropean Conference on Computer Vision. Springer, Cham, 382–400

    2024

  3. [3]

    Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. 2021. Mip-nerf: A multiscale repre- sentation for anti-aliasing neural radiance fields. InProceedings of the IEEE/CVF International Conference on Computer Vision. IEEE, Piscataway, NJ, USA, 5855– 5864

    2021

  4. [4]

    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. IEEE, Piscataway, NJ, USA, 5470–5479

    2022

  5. [5]

    Youyu Chen, Junjun Jiang, Kui Jiang, Xiao Tang, Zhihao Li, Xianming Liu, and Yinyu Nie. 2025. DashGaussian: Optimizing 3D Gaussian Splatting in 200 Seconds. InProceedings of the Computer Vision and Pattern Recognition Conference. IEEE, Piscataway, NJ, USA, 11146–11155

    2025

  6. [6]

    Guangchi Fang and Bing Wang. 2024. Mini-splatting: Representing scenes with a constrained number of gaussians. InEuropean Conference on Computer Vision. Springer, Cham, 165–181

    2024

  7. [7]

    Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A Efros, and Xiaolong Wang

  8. [8]

    Venu Madhav Govindu. 2001. Combining two-view constraints for motion esti- mation. InProceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, Vol. 2. IEEE, Piscataway, NJ, USA, II–II

    2001

  9. [9]

    Alex Hanson, Allen Tu, Geng Lin, Vasu Singla, Matthias Zwicker, and Tom Goldstein. 2025. Speedy-splat: Fast 3d gaussian splatting with sparse pixels and sparse primitives. InProceedings of the Computer Vision and Pattern Recognition Conference. IEEE, Piscataway, NJ, USA, 21537–21546

    2025

  10. [10]

    Richard Hartley, Jochen Trumpf, Yuchao Dai, and Hongdong Li. 2013. Rotation averaging.International Journal of Computer Vision103, 3 (2013), 267–305

    2013

  11. [11]

    Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. 2018. Deep blending for free-viewpoint image-based ren- dering.ACM Transactions on Graphics (ToG)37, 6 (2018), 1–15

    2018

  12. [12]

    Boyao Huang et al. 2025. 3R-GS: Efficient Gaussian Splatting without SfM Using 3D Reconstruction Model.arXiv preprint arXiv:2503.04946(2025)

    Pith/arXiv arXiv 2025

  13. [13]

    Hervé Jégou, Matthijs Douze, Cordelia Schmid, and Patrick Pérez. 2010. Aggre- gating local descriptors into a compact image representation. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition. IEEE, Piscataway, NJ, USA, 3304–3311

    2010

  14. [14]

    Zihao Ji et al. 2025. Hierarchical Tracking 3DGS: Robust SfM-Free 3D Gaussian Splatting. InProceedings of the Computer Vision and Pattern Recognition Conference. IEEE, Piscataway, NJ, USA, 6170–6180

    2025

  15. [15]

    Jeff Johnson, Matthijs Douze, and Hervé Jégou. 2019. Billion-scale similarity search with GPUs.IEEE Transactions on Big Data7, 3 (2019), 535–547

    2019

  16. [16]

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

  17. [17]

    Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Weiwei Sun, Jeff Tseng, Hossam Isber, and Andrea Tagliasacchi. 2024. 3D Gaussian Splatting as Markov Chain Monte Carlo. InAdvances in Neural Information Processing Systems, Vol. 37. Curran Associates, Inc., Red Hook, NY, USA

    2024

  18. [18]

    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

  19. [19]

    Vincent Leroy, Yohann Cabon, and Jerome Revaud. 2024. Grounding Image Match- ing in 3D with MASt3R. InEuropean Conference on Computer Vision. Springer, Cham, 71–91

    2024

  20. [20]

    Jiahao Li, Haochen Wang, Muhammad Zubair Irshad, Igor Vasiljevic, Matthew R Walter, Vitor Campagnolo Guizilini, and Greg Shakhnarovich. 2026. FastMap: Re- visiting Structure from Motion through First-Order Optimization. InProceedings of the International Conference on 3D Vision. IEEE, Piscataway, NJ, USA

    2026

  21. [21]

    Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. 2021. BARF: Bundle-Adjusting Neural Radiance Fields. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision. IEEE, Piscataway, NJ, USA, 5741–5751

    2021

  22. [22]

    Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Markus Steinberger, Francisco Vicente Carrasco, and Fernando De La Torre. 2024. Taming 3dgs: High- quality radiance fields with limited resources. InSIGGRAPH Asia 2024 Conference Papers. ACM, New York, NY, USA, 1–11

    2024

  23. [23]

    Daniel Martinec and Tomas Pajdla. 2007. Robust rotation and translation esti- mation in multiview reconstruction. In2007 IEEE Conference on Computer Vision and Pattern Recognition. IEEE, Piscataway, NJ, USA, 1–8

    2007

  24. [24]

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. 2020. Nerf: Representing scenes as neural radiance fields for view synthesis. InEuropean Conference on Computer Vision. Springer, Cham, 405–421

    2020

  25. [25]

    Pierre Moulon, Pascal Monasse, Romuald Perrot, and Renaud Marlet. 2016. Open- MVG: Open multiple view geometry. InInternational Workshop on Reproducible Research in Pattern Recognition. Springer, Cham, 60–74

    2016

  26. [26]

    Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. 2022. In- stant neural graphics primitives with a multiresolution hash encoding. InACM SIGGRAPH 2022 Conference Proceedings. ACM, New York, NY, USA, 1–15

    2022

  27. [27]

    Linfei Pan, Daniel Barath, Marc Pollefeys, and Johannes L Schönberger. 2024. Global structure-from-motion revisited. InEuropean Conference on Computer Vision. Springer, Cham, 58–77

    2024

  28. [28]

    Florent Perronnin, Jorge Sánchez, and Thomas Mensink. 2010. Improving the fisher kernel for large-scale image classification. InEuropean Conference on Computer Vision. Springer, Cham, 143–156

    2010

  29. [29]

    Shiwei Ren, Tianci Wen, Yongchun Fang, and Biao Lu. 2025. FastGS: Training 3D Gaussian Splatting in 100 Seconds. InProceedings of the Computer Vision and Pattern Recognition Conference. IEEE, Piscataway, NJ, USA, 18540–18550

    2025

  30. [30]

    Johannes L Schönberger and Jan-Michael Frahm. 2016. Structure-from-motion revisited. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition. IEEE, Piscataway, NJ, USA, 4104–4113. SalientGS: Unified SfM-to-3DGS with Importance-Guided MCMC Gaussian Allocation ACM MM ’26, November 10–14, 2026, Rio de Janeiro, Brazil

    2016

  31. [31]

    Thomas Schöps, Torsten Sattler, and Marc Pollefeys. 2019. Bad SLAM: Bundle Adjusted Direct Visual SLAM. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, Piscataway, NJ, USA, 134–144

    2019

  32. [32]

    Josef Sivic and Andrew Zisserman. 2003. Video Google: A text retrieval approach to object matching in videos. InProceedings of the IEEE International Conference on Computer Vision, Vol. 2. IEEE, Piscataway, NJ, USA, 1470–1477

    2003

  33. [33]

    Cameron Smith, Lily Goli, David Chen, Adam W Harley, and Leonidas Guibas

  34. [34]

    Chris Sweeney. 2015. Theia: A fast and scalable structure-from-motion library. InProceedings of the 23rd ACM International Conference on Multimedia. ACM, New York, NY, USA, 693–696

    2015

  35. [35]

    Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari

  36. [36]

    Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Tagliasacchi, David Novotny, Christian Rupprecht, et al. 2025. VGGT: Visual Geometry Grounded Transformer.arXiv preprint arXiv:2503.11651(2025)

    Pith/arXiv arXiv 2025

  37. [37]

    InEuropean Conference on Computer Vision

    FlowMap: High-Quality Camera Poses, Intrinsics, and Depth via Gradient Descent. InEuropean Conference on Computer Vision. Springer, Cham, 345–363

  38. [38]

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. 2024. DUSt3R: Geometric 3D Vision Made Easy. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, Piscat- away, NJ, USA, 20697–20709

    2024

  39. [39]

    Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu

  40. [40]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    SPARF: Neural radiance fields from sparse and noisy poses. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, Piscataway, NJ, USA, 4190–4200

  41. [41]

    Tianyu Xiong, Rui Li, Linjie Li, and Jiaqi Yang. 2026. GloSplat: Joint Pose- Appearance Optimization for Faster and More Accurate 3D Reconstruction.arXiv preprint arXiv:2603.04847(2026). https://arxiv.org/abs/2603.04847

    arXiv 2026

  42. [42]

    Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. 2024. VGGSfM: Visual Geometry Grounded Deep Structure From Motion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, Piscataway, NJ, USA, 21686–21697

    2024

  43. [46]

    Kyle Wilson, David Bindel, and Noah Snavely. 2020. On the distribution of minima in intrinsic-metric rotation averaging. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, Piscataway, NJ, USA, 11997–12006

    2020

  44. [48]

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang

  45. [2016]

    In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    NetVLAD: CNN architecture for weakly supervised place recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. IEEE, Piscataway, NJ, USA, 5297–5307

  46. [2018]

    InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition. IEEE, Piscataway, NJ, USA, 586–595

  47. [2021]

    NeRF–: Neural radiance fields without known camera parameters.arXiv preprint arXiv:2102.07064(2021)

    Pith/arXiv arXiv 2021

  48. [2023]

    https://repo- sam.inria.fr/fungraph/3d-gaussian-splatting/

    3D Gaussian Splatting for Real-Time Radiance Field Rendering.ACM Transactions on Graphics42, 4, Article 139 (July 2023), 14 pages. https://repo- sam.inria.fr/fungraph/3d-gaussian-splatting/

    2023

  49. [2024]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    COLMAP-Free 3D Gaussian Splatting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, Piscataway, NJ, USA, 20796–20805