Pith. sign in

REVIEW 3 major objections 5 minor 33 references

G2SR claims that given cross-view 2D splat correspondences, 3D splats follow analytically from multi-view geometry, enabling Gaussian-based surface reconstruction that matches or exceeds state-of-the-art end-to-end methods while running at

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 01:58 UTC pith:WH6R6PWU

load-bearing objection A genuinely new architecture for fast sparse-view Gaussian reconstruction, with real speed/memory wins, but the headline accuracy claim is weakened by a valid-pixel mask that lets it skip the hardest quarter of the image. the 3 major comments →

arxiv 2607.14470 v1 pith:WH6R6PWU submitted 2026-07-16 cs.CV cs.RO

G²SR: Geometric Methods for Fast and Memory-Efficient Gaussian-based Surface Reconstruction

Dasong Gao , Vivienne Sze , Sertac Karaman This is my paper
classification cs.CV cs.RO
keywords surface reconstruction3D Gaussian splattingfew-view reconstructionmulti-view geometrytriangulationoptical flowmemory efficiencyonline mapping
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.

The paper proposes G2SR, a framework for few-view surface reconstruction that splits the task into a lightweight neural frontend and an analytic backend. The frontend detects 2D Gaussian splats on a reference image and tracks them across views using optical flow; the backend triangulates each splat into a metric-scale 3D surface patch using classical multi-view geometry. The central claim is that the ill-posed part of sparse-view reconstruction is only the image-plane detection and tracking, while the 3D structure itself is well-posed once correspondences are known. On ScanNet, Replica, and DTU, G2SR reports lower depth error and comparable or better mesh accuracy than heavier end-to-end networks, while using 5–107x less GPU memory and achieving up to 18x higher throughput. A sympathetic reader would see this as evidence that hybrid analytic-neural pipelines can make online 3D reconstruction practical on resource-constrained robots.

Core claim

G2SR establishes that a Gaussian splat lying on a scene surface is fully determined by its 2D projections across input views, so once those projections - the 2D splats - are known, the 3D splat's position, orientation, and scale can be recovered analytically by triangulation. The paper shows that this projection-matching formulation can be solved with a Gauss-Newton optimizer that minimizes the squared Hellinger distance between each splat's projection and its observed 2D target, initialized by DLT for the center and by affine-map normal estimation for the orientation. Because the analytic backend replaces the need for a large network to regress 3D structure, the neural frontend can be tiny

What carries the argument

The central mechanism is the detect-track-triangulate pipeline, mirroring feature-based SLAM. A lightweight convolutional network with iterative in-block cross-attention detects boundary-respecting 2D Gaussian splats on the reference image, trained with a rendering loss under random permutation to enforce order-robustness. An affine tracking algorithm estimates correspondences across views by warping each 2D splat's sigma points with optical flow, solving a closed-form 4x4 least-squares problem for the affine map, and rejecting inconsistent matches with forward-backward flow. A geometric triangulator then recovers each 3D splat by minimizing the squared Hellinger distance between its project

Load-bearing premise

The pipeline's accuracy rests on the assumption that the optical flow and the boundary-respecting 2D splat detector provide correct correspondences for the same physical surface points across views, even near occlusions, textureless regions, and lighting changes.

What would settle it

Construct scenes with known ground-truth surfaces containing large textureless or occluded regions, run G2SR, and check whether reconstructed depth diverges from ground truth beyond the reported error; if a single textureless wall or a foreground occluder causes dense, plausible-but-wrong splats that pass the forward-backward check, the correspondence assumption is falsified.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the claim holds, few-view Gaussian-based reconstruction becomes fast and memory-light enough to run online on drones and wearables, enabling real-time mapping during exploration.
  • Because triangulation recovers metric scale from camera poses, downstream tasks like navigation and manipulation no longer need post-hoc alignment to world scale.
  • The decoupling suggests that further improvements in detection and tracking networks will directly translate into better geometry, without requiring larger 3D decoding networks.
  • The approach's robustness to camera intrinsics and view configuration shifts implies it can generalize across datasets without per-scene tuning, a practical advantage for deployment.
  • The reported geometric accuracy, with depth error roughly half that of the strongest baseline on Replica 2-view, indicates that analytic geometry can outperform learned depth regression in sparse settings.

Where Pith is reading between the lines

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

  • The paper's framing suggests a general recipe: identify the well-posed core of an ill-posed reconstruction problem and solve it analytically, using learning only for the genuinely ambiguous subproblems. This recipe may extend beyond Gaussian splats to other primitive-based representations like surfels or planar patches.
  • Because the frontend is trained only on RealEstate10K and the optical flow network is used off-the-shelf, the bottleneck may shift to correspondence quality under occlusion or textureless regions. One testable extension is to add explicit occlusion-aware weighting when aggregating multi-view flow signals.
  • The aggressive outlier rejection limits coverage; accumulating splats over consecutive frames (as hinted by the conclusion) could both recover missing boundary regions and improve accuracy, potentially enabling a fully online SLAM-style Gaussian mapping system.
  • The Hellinger-based projection matching with Gauss-Newton iterations is a generic alignment objective; it could be repurposed for joint refinement of splats from many views or for aligning predicted 2D splats to multi-frame video streams.

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

3 major / 5 minor

Summary. The paper proposes G2SR, a feed-forward few-view surface reconstruction method that decouples the task into a lightweight neural frontend (2D Gaussian splat detection plus optical-flow-based affine tracking) and an analytic backend that triangulates the detected 2D splats into metric-scale 3D Gaussian splats. The authors claim that on Replica, ScanNet, and DTU, G2SR matches or exceeds the geometric accuracy of end-to-end baselines (MVSplat, FreeSplat, MonoSplat, C3G, SurfelSplat) while using 115--203 MB of GPU memory and running at 69--89 reconstructions per second. They further claim that the geometric backend makes the method robust to camera-geometry shifts and that G2SR produces metric-scale depth without median alignment to ground truth.

Significance. If the claims hold, the detect-track-triangulate architecture is a meaningful contribution: it replaces a large part of learned 3D regression with classical multi-view geometry, which is parameter-efficient and principled. The paper gives a concrete formulation (Sections III-A--D), evaluates on standard datasets against several strong baselines, and explicitly reports a coverage gap and rendering trade-offs. The proposed method's efficiency advantage and the decision to evaluate G2SR without the median alignment granted to baselines are notable strengths. However, the central accuracy claim is currently supported only under a per-method valid-pixel mask, and no code is released, so the result is not fully independently verifiable.

major comments (3)
  1. [§IV-A2, Tables I and III] The headline accuracy comparison is evaluated over each method's own valid-pixel set. The text states that all methods are evaluated over valid pixels with rendered opacity α≥0.5. Table III reports G2SR coverage of 85.8% (Replica) and 73.1% (ScanNet) for 3-view inputs, while every baseline exceeds 94%. Consequently, Table I's Abs Rel of 7.7% on ScanNet 3-view is computed after excluding 26.9% of pixels. Section IV-D attributes this exclusion to aggressive outlier rejection at boundaries and occlusions. Thus the abstract's claim that G2SR 'matches or exceeds the geometric accuracy of state-of-the-art methods' is only established on the subset of pixels G2SR chooses to reconstruct. Please recompute depth metrics on the full image or on a common mask, and report coverage in Table I; otherwise the central accuracy claim is overstated.
  2. [§IV-A2, §IV-C1] Tables I and II report point estimates without any variance, and the number of test scenes per dataset is not stated. The paper also does not mention code or model release. Given the claim that G2SR 'matches or exceeds' accuracy across datasets and methods, and given the large claimed efficiency margins, per-scene results or error bars are needed to assess robustness. A statement of code/data availability would also allow independent verification of the memory and throughput measurements.
  3. [§III-C, §IV-D] The method's accuracy rests on the correctness of NeuFlowV2 correspondences and the detector's 2D splats, but the paper offers no direct evaluation of correspondence quality. The forward-backward check in Section III-C rejects only translation-inconsistent splats; affine errors in A can survive and bias the triangulated splat's orientation and scale. The paper instead reports coverage as an afterthought and explains it as 'aggressive outlier rejection' (§IV-D). A threshold sweep and an ablation of the forward-backward outlier rejection on both accuracy and coverage, plus a comparison of surviving vs. rejected splats, would clarify whether the reported gains come from geometry or from discarding hard pixels. Without this, the empirical bottleneck of the method is unmeasured.
minor comments (5)
  1. [§V] The conclusion says 'resolution>384×512' while the abstract and experiments use '384×512'; please correct the inconsistency.
  2. [§III-D] The statement that two views supply 2×5=10 DoF of constraints is imprecise, since the two projections are not independent under epipolar geometry. Rephrase as 'ten scalar measurements' or similar.
  3. [Figure 3] Because the quantitative metrics exclude pixels with rendered opacity below 0.5, the qualitative figure should include a coverage/valid-pixel mask panel so the reader can see which regions are excluded for G2SR and for the baselines.
  4. [Table II] C3G is listed as '-' for CD(T). Please state explicitly that it could not be evaluated under the TSDF protocol, or provide the reason.
  5. [§III-D] The term 'analytic backend' may be misleading because the Gauss-Newton solver is iterative rather than closed-form. Consider calling it a 'geometric backend' or clarifying the distinction.

Circularity Check

0 steps flagged

No significant circularity: 3D splats are triangulated from measured 2D correspondences with no target-fitting; self-citations are non-load-bearing.

full rationale

The central derivation chain is not circular. G2SR's 3D splats are obtained by solving the triangulation/projection-matching problem in Eq. (7) from 2D splat correspondences, which are produced by a lightweight detector and pretrained NeuFlowV2 optical flow, not from the 3D result or from evaluation ground truth. The detection network is trained on RealEstate10K with the rendering loss of Eq. (3), and the paper explicitly states: 'We only train the splat detection network Φ (Section III-B) on RealEstate10K [28] and use the pretrained NeuFlowV2 [29] optical flow network off-the-shelf. No fine-tuning is performed on Replica, ScanNet, or DTU.' Evaluation does not fit G2SR to the target: 'All methods except G2SR are median-aligned to metric scale using ground truth depth before evaluating,' while G2SR is evaluated unaligned. The two self-citations in the paper ([3], [11]) support background statements about DRAM energy and floater artifacts, respectively; neither is load-bearing for the claimed analytical triangulation or the accuracy results. The valid-pixel/coverage issue raised in Section IV-A2 and Table III is a disclosed evaluation-scope limitation, not a circular derivation: the depth metrics are computed only over pixels the method chooses to reconstruct, and the paper reports the resulting lower coverage (85.8% / 73.1% for 3-view Replica/ScanNet) alongside the accuracy numbers. That affects how the headline accuracy claim should be interpreted, but it does not make the derived 3D splats equivalent to the evaluation inputs or to any fitted parameter. Therefore no circular step is present.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The central geometric claim rests on standard multi-view geometry and a small set of domain assumptions about surface representability and correspondence quality. No new physical entities are introduced; the learned detector and flow network are the main empirical inputs, and their training hyperparameters are not fully specified.

free parameters (6)
  • Detector network Φ weights = 0.41M parameters, trained on RealEstate10K
    Learned 2D splat detector; its predictions determine triangulation quality. Fitted to data, not to evaluation depth.
  • Splat thickness ε = fixed small positive (value not specified)
    Assumes planar surfels; chosen by hand in Section III-A.
  • Forward-backward consistency threshold = 3 pixels
    Hard threshold for outlier rejection; affects coverage and accuracy (Section III-C).
  • Gauss-Newton iterations / step size = N=20, η not specified
    Optimization hyperparameters controlling convergence of triangulation (Section III-D).
  • Detector training loss weights λ1, λ2 = not reported
    Weights balancing L1 and SSIM losses in Eq (3); central to detector training.
  • Detector block size s and iterations T = not reported
    Control number of splats; architecture choices affecting reconstruction detail.
axioms (4)
  • domain assumption Camera poses and intrinsics are known and accurate for input views.
    Triangulation in Section III-D uses given extrinsics/intrinsics; metric scale and geometry rest on them.
  • domain assumption Detected 2D splats and optical-flow correspondences track the same physical surface points across views.
    Tracking process (Section III-C) assumes correspondences; incorrect correspondences produce wrong triangulated splats.
  • domain assumption Scene surfaces are locally planar and can be represented by thin elliptical Gaussian disks with negligible thickness.
    Splat geometry G=(μ,Σ) with fixed ε in Section III-A; all triangulation assumes this representation.
  • domain assumption The rendering loss on 2D splats, trained with random permutation, yields an ordering-invariant set that covers surfaces and respects boundaries.
    Training discussion in Section III-B justifies using 2D splats as projection targets.

pith-pipeline@v1.3.0-alltime-deepseek · 14627 in / 12356 out tokens · 106877 ms · 2026-08-02T01:58:52.897660+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of G$^2$SR: Geometric Methods for Fast and Memory-Efficient Gaussian-based Surface Reconstruction." pith.science (2026). https://pith.science/paper/WH6R6PWU

@misc{pith2026260714470,
  author       = {Pith},
  title        = {Pith review of: G$^2$SR: Geometric Methods for Fast and Memory-Efficient Gaussian-based Surface Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WH6R6PWU}},
  note         = {Machine review of arXiv:2607.14470}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Few-view surface reconstruction recovers the visible surfaces of a scene from a few posed RGB images, providing the 3D models that robots need to explore and interact online. On mobile platforms, the reconstruction must be fast and geometrically accurate while keeping a small memory footprint to ensure safe and efficient operation. 3D Gaussian Splatting (3DGS) offers a high-fidelity scene representation, but building it from a few views is ill-posed, as many distinct surfaces reproduce the same images, making traditional photometric methods prone to "floater" artifacts. End-to-end methods resolve the ambiguity by regressing splats with large, usually Transformer-based, networks that require heavy compute and memory while generalizing poorly to new scenes. We propose G2SR, which exploits a well-posed core of the task: given cross-view 2D splat correspondences, 3D splats follow analytically from multi-view geometry. G2SR employs a lightweight neural frontend to detect and track 2D Gaussian splats on the image plane and an analytic backend to triangulate each into a metric-scale 3D splat. On ScanNet, Replica, and DTU, G2SR matches or exceeds the geometric accuracy of state-of-the-art end-to-end methods while running at 69-89 reconstructions per second within 203 MB of GPU memory (5-107x less) for 2- and 3-view inputs at 384 x 512 resolution, offering a practical path to online Gaussian-based surface reconstruction.

Figures

Figures reproduced from arXiv: 2607.14470 by Dasong Gao, Sertac Karaman, Vivienne Sze.

Figure 1
Figure 1. Figure 1: From a few (2–3) posed RGB images, prior end-to-end methods (top) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the G2SR pipeline. A lightweight network detects 2D Gaussian splats {G¯r i } on the reference image Ir (Section III-B); an optical-flow￾based sigma-point warping then establishes affine correspondences {G¯k i } to view k (Section III-C); finally, a Gauss–Newton solver triangulates 3D splats {Gˆ i} by matching projections under the squared Hellinger distance d 2 H (Section III-D). During 3-view … view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison on a Replica 2-view reconstruction. Compared with baselines, G [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

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

33 extracted references · 5 linked inside Pith

  1. [1]

    The replica dataset: A digital replica of indoor spaces,

    J. Straub, T. Whelan, L. Ma, Y . Chen, E. Wijmans, S. Green, J. J. Engel, R. Mur-Artal, C. Ren, S. Vermaet al., “The replica dataset: A digital replica of indoor spaces,”arXiv preprint arXiv:1906.05797, 2019

    Pith/arXiv arXiv 1906

  2. [2]

    Scannet: Richly-annotated 3d reconstructions of indoor scenes,

    A. Dai, A. X. Chang, M. Savva, M. Halber, T. Funkhouser, and M. Nießner, “Scannet: Richly-annotated 3d reconstructions of indoor scenes,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 5828–5839

    2017

  3. [3]

    Gmmap: Memory-efficient contin- uous occupancy map using gaussian mixture model,

    P. Z. X. Li, S. Karaman, and V . Sze, “Gmmap: Memory-efficient contin- uous occupancy map using gaussian mixture model,”IEEE Transactions on Robotics, vol. 40, pp. 1339–1355, 2024

    2024

  4. [4]

    Efficientvit: Memory efficient vision transformer with cascaded group attention,

    X. Liu, H. Peng, N. Zheng, Y . Yang, H. Hu, and Y . Yuan, “Efficientvit: Memory efficient vision transformer with cascaded group attention,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 14 420–14 430

    2023

  5. [5]

    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 Transactions on Graphics (ToG), vol. 42, no. 4, pp. 1–14, 2023

    2023

  6. [6]

    2d gaussian splatting for geometrically accurate radiance fields,

    B. Huang, Z. Yu, A. Chen, A. Geiger, and S. Gao, “2d gaussian splatting for geometrically accurate radiance fields,” inACM SIGGRAPH 2024 conference papers, 2024, pp. 1–11

    2024

  7. [7]

    Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering,

    A. Gu ´edon and V . Lepetit, “Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 5354–5363

    2024

  8. [8]

    Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction,

    D. Chen, H. Li, W. Ye, Y . Wang, W. Xie, S. Zhai, N. Wang, H. Liu, H. Bao, and G. Zhang, “Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction,”IEEE Transactions on Visualization and Computer Graphics, vol. 31, no. 9, pp. 6100–6111, 2024

    2024

  9. [9]

    Sparse2dgs: Geometry-prioritized gaussian splatting for surface reconstruction from sparse views,

    J. Wu, R. Li, Y . Zhu, R. Guo, J. Sun, and Y . Zhang, “Sparse2dgs: Geometry-prioritized gaussian splatting for surface reconstruction from sparse views,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 11 307–11 316

    2025

  10. [10]

    Fatesgs: Fast and accurate sparse-view surface reconstruction using gaussian splatting with depth-feature consistency,

    H. Huang, Y . Wu, C. Deng, G. Gao, M. Gu, and Y .-S. Liu, “Fatesgs: Fast and accurate sparse-view surface reconstruction using gaussian splatting with depth-feature consistency,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 4, 2025, pp. 3644–3652

    2025

  11. [11]

    Gevo: Memory- efficient monocular visual odometry using gaussians,

    D. Gao, P. Z. X. Li, V . Sze, and S. Karaman, “Gevo: Memory- efficient monocular visual odometry using gaussians,”IEEE Robotics and Automation Letters, 2025

    2025

  12. [12]

    Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images,

    Y . Chen, H. Xu, C. Zheng, B. Zhuang, M. Pollefeys, A. Geiger, T.- J. Cham, and J. Cai, “Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images,” inEuropean conference on computer vision. Springer, 2024, pp. 370–386

    2024

  13. [13]

    Freesplat: Generalizable 3d gaussian splatting towards free view synthesis of indoor scenes,

    Y . Wang, T. Huang, H. Chen, and G. H. Lee, “Freesplat: Generalizable 3d gaussian splatting towards free view synthesis of indoor scenes,”Ad- vances in Neural Information Processing Systems, vol. 37, pp. 107 326– 107 349, 2024

    2024

  14. [14]

    Monosplat: Generalizable 3d gaussian splatting from monocular depth foundation models,

    Y . Liu, K. Fan, W. Yu, C. Li, H. Lu, and Y . Yuan, “Monosplat: Generalizable 3d gaussian splatting from monocular depth foundation models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 21 570–21 579

    2025

  15. [15]

    C3g: Learning compact 3d representations with 2k gaussians,

    H. An, J. Jung, M. Kim, S. Hong, C. Kim, K. Fukuda, M. Jeon, J. Han, T. Narihira, H. Koet al., “C3g: Learning compact 3d representations with 2k gaussians,”arXiv preprint arXiv:2512.04021, 2025

    Pith/arXiv arXiv 2025

  16. [16]

    Surfelsplat: Learning efficient and generalizable gaussian surfel representations for sparse- view surface reconstruction,

    C. Dai, S. Zhang, M. Chen, and Y . Duan, “Surfelsplat: Learning efficient and generalizable gaussian surfel representations for sparse- view surface reconstruction,” inThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    2025

  17. [17]

    Large scale multi-view stereopsis evaluation,

    R. Jensen, A. Dahl, G. V ogiatzis, E. Tola, and H. Aanæs, “Large scale multi-view stereopsis evaluation,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2014, pp. 406–413

    2014

  18. [18]

    Mvsgaussian: Fast generalizable gaussian splatting recon- struction from multi-view stereo,

    T. Liu, G. Wang, S. Hu, L. Shen, X. Ye, Y . Zang, Z. Cao, W. Li, and Z. Liu, “Mvsgaussian: Fast generalizable gaussian splatting recon- struction from multi-view stereo,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 37–53

    2024

  19. [19]

    Depthsplat: Connecting gaussian splatting and depth,

    H. Xu, S. Peng, F. Wang, H. Blum, D. Barath, A. Geiger, and M. Polle- feys, “Depthsplat: Connecting gaussian splatting and depth,” inProceed- ings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 16 453–16 463

    2025

  20. [20]

    Mugs: Multi-baseline generalizable gaussian splatting reconstruction,

    Y . Lou, L. Shen, T. Liu, J. Li, Z. Huang, H. Sun, and Z. Cao, “Mugs: Multi-baseline generalizable gaussian splatting reconstruction,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 25 583–25 593

    2025

  21. [21]

    Transplat: Generalizable 3d gaussian splatting from sparse multi-view images with transformers,

    C. Zhang, Y . Zou, Z. Li, M. Yi, and H. Wang, “Transplat: Generalizable 3d gaussian splatting from sparse multi-view images with transformers,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 9, 2025, pp. 9869–9877

    2025

  22. [22]

    G3splat: Geometrically consistent generalizable gaussian splatting,

    M. Hosseinzadeh, S.-F. Chng, Y . Xu, S. Lucey, I. Reid, and R. Garg, “G3splat: Geometrically consistent generalizable gaussian splatting,” arXiv preprint arXiv:2512.17547, 2025

    Pith/arXiv arXiv 2025

  23. [23]

    Off the grid: Detection of primitives for feed- forward 3d gaussian splatting,

    A. Moreau, R. Shaw, M. Nazarczuk, J. Shin, T. Tanay, Z. Zhang, S. Xu, and E. P ´erez-Pellitero, “Off the grid: Detection of primitives for feed- forward 3d gaussian splatting,”arXiv preprint arXiv:2512.15508, 2025

    arXiv 2025

  24. [24]

    pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d re- construction,

    D. Charatan, S. L. Li, A. Tagliasacchi, and V . Sitzmann, “pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d re- construction,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 19 457–19 467

    2024

  25. [25]

    Edgs: Eliminat- ing densification for efficient convergence of 3dgs,

    D. Kotovenko, O. Grebenkova, and B. Ommer, “Edgs: Eliminat- ing densification for efficient convergence of 3dgs,”arXiv preprint arXiv:2504.13204, 2025

    arXiv 2025

  26. [26]

    Triags: Differentiable triangulation-guided geometric consistency for 3d gaussian splatting,

    Q. Tran and T. Dang, “Triags: Differentiable triangulation-guided geometric consistency for 3d gaussian splatting,”arXiv preprint arXiv:2512.06269, 2025

    arXiv 2025

  27. [27]

    Fsfsplatter: Build surface and novel views with sparse-views within 2min,

    Y . Zhao, Y . Pan, J. Nan, L. Chen, and J. Yi, “Fsfsplatter: Build surface and novel views with sparse-views within 2min,”arXiv preprint arXiv:2510.02691, 2025

    arXiv 2025

  28. [28]

    Stereo magnification: Learning view synthesis using multiplane images,

    T. Zhou, R. Tucker, J. Flynn, G. Fyffe, and N. Snavely, “Stereo magnification: Learning view synthesis using multiplane images,”arXiv preprint arXiv:1805.09817, 2018

    Pith/arXiv arXiv 2018

  29. [29]

    Neuflow-v2: Push high- efficiency optical flow to the limit,

    Z. Zhang, A. Gupta, H. Jiang, and H. Singh, “Neuflow-v2: Push high- efficiency optical flow to the limit,” in2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2025, pp. 2479–2485

    2025

  30. [30]

    Dynamic map building and localization: New theo- retical foundations,

    J. K. Uhlmann, “Dynamic map building and localization: New theo- retical foundations,” Ph.D. dissertation, University of Oxford Oxford, 1995

    1995

  31. [31]

    Triangulation,

    R. I. Hartley and P. Sturm, “Triangulation,”Computer vision and image understanding, vol. 68, no. 2, pp. 146–157, 1997

    1997

  32. [32]

    Computer vision meets geometric modeling: Multi-view reconstruction of surface points and normals using affine correspondences,

    I. Eichhardt and L. Hajder, “Computer vision meets geometric modeling: Multi-view reconstruction of surface points and normals using affine correspondences,” inProceedings of the IEEE International Conference on Computer Vision Workshops, 2017, pp. 2427–2435

    2017

  33. [33]

    Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction,

    P. Wang, L. Liu, Y . Liu, C. Theobalt, T. Komura, and W. Wang, “Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction,”arXiv preprint arXiv:2106.10689, 2021

    Pith/arXiv arXiv 2021