Pith. sign in

REVIEW 3 major objections 5 minor 153 references

Immediate 3D Gaussian Splat Reconstruction of Unordered Input with Global Consistency

T0 review · 3 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read This paper claims that unordered, out-of-order photo captures can be turned into live, globally consistent 3D Gaussian Splat reconstructions immediately, without waiting for an offline structure-from-motion pass.

desk verdict This is a stronger paper than the stress-test note implies; the unmeasured retrieval recall is a real reporting gap, but the end-to-end evaluation supports the central claim well enough to warrant a serious referee. read the letter →

arxiv 2607.14481 v1 pith:PPHWO6AP submitted 2026-07-16 cs.CV

Andreas Meuleman , Linus Franke , Boris Zhestiankin , Camille Montemagni , George Drettakis This is my paper · ORCID
classification cs.CV
keywords 3DGaussianSplattingunorderedinputimmediatefeedbackloopclosurecovisibilitygraphvisualplacerecognitionprogressivehierarchyonlinereconstruction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper set out to solve a practical problem: 3D Gaussian Splatting captures usually mix continuous video with out-of-order stills, yet existing online reconstruction assumes ordered input and offline structure-from-motion makes you wait. The authors claim the first immediate-feedback solution that keeps global consistency for such unordered captures. They achieve it by replacing nearest-in-time matching with a two-stage retrieval-and-verification cascade, using a weighted covisibility graph to find well-connected keyframes, and adding cluster-based loop closure with a fast graph-propagation update. If it holds, a user can walk around a scene, turn, revisit, and reshoot, and see a coherent live reconstruction within about a second, with quality within roughly 1-2 dB of an offline global reconstruction at a fraction of the compute time.

What carries the argument

The central object is the weighted covisibility graph: a graph whose nodes are keyframes and whose edge weights are inverse inlier counts from verified matches. It performs three jobs: it lets a greedy traversal select the 20 most strongly connected keyframes for local bundle adjustment, it exposes disjoint clusters that indicate revisits in unordered input, and it provides a Dijkstra backbone along which the loop-closure similarity correction is interpolated and propagated to all poses and Gaussian primitives. Around this graph, the paper builds a speed/quality cascade: a global descriptor quickly retrieves 20 candidates, a fast local feature matcher cuts that to five, and a slower accurate

What would settle it

Run the pipeline on a scene with repeated, textureless regions or sudden lighting changes and count how many frames are shelved because none of the top-20 global-descriptor matches passes the 500-inlier geometric verification threshold; then compare the final novel-view PSNR against the same frames processed offline by a global structure-from-motion pipeline. If shelving is frequent or the quality gap widens sharply, the global-consistency promise breaks.

Watch

Extended reading notes

Core claim

The paper claims that a radiance-field capture workflow no longer has to choose between immediacy and global consistency. It introduces a matching cascade in which a global image descriptor quickly finds 20 candidate keyframes for each incoming frame, a fast local matcher narrows those to five, and a slower accurate matcher verifies which pairs truly overlap; a weighted covisibility graph then lets the system select the 20 most connected keyframes for bundle adjustment, regardless of when they were captured. When the graph reveals two disconnected clusters that nevertheless see the same content, a cluster-based welding window re-optimizes poses together, and the resulting similarity correcti

Load-bearing premise

The whole pipeline rests on the hope that for every frame that truly overlaps the scene, at least one of the 20 most similar global descriptors will survive the geometric verification step; if that fails, the frame is shelved and the scene silently loses coverage.

Editorial extensions

If this is right

  • Users can capture scenes by moving freely, turning, and revisiting areas, and see a live reconstruction instead of waiting for a full offline structure-from-motion pass.
  • The method handles unordered and shuffled input with only a small quality loss: on the MipNeRF360 scenes, random shuffling drops PSNR from 26.29 to 25.73.
  • On large scenes with thousands of images, the progressive hierarchy keeps peak trained Gaussians and GPU memory bounded; on a city-scale scene, it stays below 3.7 million Gaussians and 12.8 GB where the same run without the hierarchy runs out of memory before 20% of frames.
  • Quality is within roughly 1-2 dB of an offline global reference while taking minutes instead of tens of minutes or hours, making immediate feedback practical within a single session.
  • Even for ordered captures, the graph-based keyframe selection improves quality over prior online methods because it repeatedly reaches far back in time, not just the last few frames.

Reading between the lines

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

  • The matching cascade is a modular speed/quality funnel, so its components could be swapped for stronger or lighter alternatives depending on hardware, scene type, or capture mode; the paper's specific choices are not intrinsic to the idea.
  • The shelving mechanism implies an implicit completeness guarantee: final quality depends on whether shelved frames are eventually re-matched, so a capture-time indicator showing the fraction of shelved frames could serve as a practical completeness meter.
  • Because the cluster-based loop closure avoids timestamps entirely, the same approach could generalize to multi-session or crowd-sourced captures where images arrive in no particular order.
  • The progressive hierarchy's merging and retrieval rules could double as a compression or streaming mechanism, potentially enabling progressive delivery of large scenes to viewers with limited memory.
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

3 major / 5 minor

Summary. This paper proposes an online 3D Gaussian Splatting reconstruction system that accepts an unordered stream of RGB images and produces a globally consistent radiance field with immediate feedback. The pipeline combines a two-level matching cascade (MixVPR retrieval, XFeat geometric preselection, LightGlue+RANSAC verification), a weighted covisibility graph for retrieving well-connected keyframe subsets, GPU local bundle adjustment with 20 keyframes, per-tile monocular-depth alignment for Gaussian placement, cluster-based loop detection and a Sim(3) graph-propagation drift correction, and a progressive hierarchy for large scenes. The evaluation compares against six pose-free baselines and an offline GLOMAP+Taming3DGS reference on TUM, MipNeRF360, StaticHikes, Tanks and Temples, Deep Blending, and large-scale CityWalk/Wave/SmallCity, reporting quality close to the offline reference at a fraction of the compute, with ablations and hyperparameter sensitivity.

Significance. If the claims hold, this is a substantial practical advance: it removes the ordered-capture restriction for immediate-feedback 3DGS and provides loop closure without timestamps. The evaluation is broad (8+ datasets), uses external baselines and an offline reference, includes ablations and latency measurements, and honestly declares limitations (quality gap, pure rotation, dynamic scenes). The main weakness is evidentiary: the reliability of the matching cascade and the quantitative validation of global consistency on unordered input are not established, and these are exactly the load-bearing components of the central claim. The paper does not ship code, but the experimental tables are detailed enough to reproduce the main trends.

major comments (3)
  1. [Sec. 4.1, Sec. 5.3, App. B.2] The completeness claim requires that every frame with true overlap survives the cascade MixVPR top-20 -> XFeat top-5 -> LightGlue+RANSAC (tau_m=500). Frames that do not are shelved (Sec. 5.3) and retried only if global-descriptor similarity grows. No per-dataset recall@20, cascade survival, shelf count, unshelve rate, or final unregistered fraction is reported. App. B.2 reports only top-5 overlap (100% MipNeRF360; 92% on a synthetic repetition scene), which is not recall@20 and implies 8% of true relevant frames fall outside top-5. This unmeasured false-negative rate directly limits global consistency. Please report these statistics and stress-test textureless/repetitive scenes.
  2. [Sec. 6 vs. Tabs. 2/4] Global consistency on unordered input is never directly measured. Pose errors (Tab. 4) and ablations (Tabs. 6/14) use ordered TUM/MipNeRF360 scenes; Tab. 2 for unordered data reports only NVS metrics. Loop closure (Sec. 6) is not evaluated with trajectory errors before/after welding, loop-closure precision/recall, or comparison of propagated poses to the GLOMAP reference on shuffled sequences. Without these, the headline claim is supported only by rendering quality. Please add ATE/RPE on shuffled MipNeRF360/Tanks/Deep Blending before/after loop closure and report the registered-keyframe fraction.
  3. [Sec. 4.1 and Tab. 12] The cascade parameters K=20, K'=5, tau_m=500 are load-bearing free parameters, but the sensitivity study varies only tau_lc and tau_v. Since a false negative at any stage permanently shelves a frame, the paper should ablate K, K', and tau_m (at least on one unordered dataset) to show the central result is not tuned to a narrow operating point.
minor comments (5)
  1. [Sec. 1] The claim 'first immediate feedback solution' should be qualified with 'to our knowledge' given the fast-moving literature.
  2. [Fig. 11] Describe how the synthetic repetition scene was generated and define 'overlap' precisely.
  3. [Table 5] Report multiple shuffle seeds rather than a single shuffled run to assess variance.
  4. [Sec. 7.2] The threshold tau_h = 1 is introduced without units; specify whether it is screen size in pixels or another quantity.
  5. [Table 3] The asterisks on SmallCity* and Wave* are unexplained; add a note.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation found; minor self-citation for protocol/baselines is not load-bearing.

full rationale

The paper's load-bearing claims—immediate feedback, unordered-input matching, global consistency via loop closure, and scalability via the progressive hierarchy—are empirical system claims evaluated against external benchmarks (TUM, MipNeRF360, Tanks and Temples, Deep Blending, CityWalk) and external baselines (Photo-SLAM, MonoGS, LongSplat, S3PO-GS, GLOMAP+Taming3DGS). The matching cascade (Sec. 4.1) does not derive its outputs from its inputs: MixVPR top-K, XFeat top-K', LightGlue+RANSAC tau_m=500 define candidate selection and verification, but recall/false-negative behavior is an experimental property, not a tautology. The covisibility graph and cluster-based loop detection (Sec. 6) are algorithmic constructions whose quality is measured via PSNR/SSIM/LPIPS and pose errors, not assumed. Gaussian placement and hierarchy performance are likewise benchmarked (peak memory, time, quality). The main self-reference is using the dataset/evaluation protocol and baseline 'On-The-Fly NVS' from Meuleman et al. 2025; this is a legitimate baseline, and the central quality results are independently grounded by external datasets and by comparison with GLOMAP/Taming3DGS. No equation is shown to reduce to a fitted constant or to the target claim. The reviewer concern about unmeasured retrieval recall and frame shelving (Sec. 5.3, Appendix B.2) is a correctness/completeness risk, not circularity: a false-negative in the cascade would weaken the global-consistency claim empirically but would not make the claim true by definition. Score 2 reflects the minor self-citation, not load-bearing circularity.

Assumptions & free parameters 12 free parameters · 4 assumptions · 0 invented entities

The central claim is an engineering feasibility claim, so the ledger is dominated by domain assumptions and hand-chosen thresholds. The matching cascade introduces no new physical entities. Most thresholds have been only partly sensitivity-tested (tau_lc and tau_v in Tab. 12; other parameters untested), so the reliability of the method outside the tested datasets rests on unquantified assumptions.

free parameters (12)
  • K (top MixVPR candidates) = 20
    Number of VPR-retrieved candidates per query; hand-chosen, not sensitivity-analyzed.
  • K' (geometric verification candidates) = 5
    Number of XFeat-filtered candidates retained for LightGlue verification.
  • tau_m (inlier threshold) = 500
    LightGlue inliers after RANSAC required to validate a keyframe pair; core gate for matching and graph edges.
  • tau_lc (loop-closure hop threshold) = 5
    Clusters separated by more than 5 covisibility hops trigger loop closure; sensitivity reported in Tab. 12.
  • tau_v (minimum cluster pairs) = 2
    Loop closure requires the smaller cluster to have more than 2 validated pairs; sensitivity reported in Tab. 12.
  • N (local BA keyframes) = 20
    Window size for local bundle adjustment; larger than prior methods; hand-set.
  • K_landmarks (BA sampling) = 10000
    Fixed number of landmarks sampled for GPU bundle adjustment; described as 'in our tests'.
  • n_f (fixed keyframes in BA) = 4
    Number of randomly chosen keyframes fixed during local BA.
  • tau_l (subpixel merge threshold) = 0.5 px
    Gaussians with screen size below 0.5 px are merged; merging triggered when >20% qualify.
  • tau_h (hierarchy reload threshold) = 1 px
    Coarse node reloaded when its contribution exceeds 1 px screen size.
  • Per-frame Gaussian optimization budget = 30 iterations
    Fixed optimization iterations per keyframe for reproducible evaluation; async until next keyframe in live capture.
  • Per-tile depth scale and offset = per-tile, per-frame
    Estimated per-tile alignment of Depth Anything 3 predicted depth to BA landmarks; NoLoDepAlgn ablation shows the per-tile version improves quality.
assumptions (4)
  • domain assumption Monocular depth from Depth Anything 3 provides a geometrically usable prior for pose and landmark initialization, and for Gaussian placement after per-tile scale-offset alignment.
    Sec. 5 bootstrap; correctness of subsequent geometric estimation relies on this learned prior being approximately scale-consistent across tiles.
  • domain assumption The covisibility graph edge weight w_ij = 1/#inliers is a sufficient proxy for geometric confidence for greedy keyframe selection, loop clustering, and Dijkstra backbone propagation.
    Sec. 4.2 and 6.3; all local BA and loop-closure propagation quality rests on inlier count correlating with pose reliability.
  • domain assumption A single Sim(3) transformation interpolated along the backbone path corrects drift for all poses and attached Gaussians, with no spatially non-uniform deformation needed.
    Sec. 6.3, Eqs. 4-9; the NoGraphProp ablation supports parity with pose-graph optimization, but only on tested scenes.
  • standard math Standard bundle adjustment and SE(3)/Sim(3) manifold optimization machinery.
    Sec. 5.1 and Appendix A.3; unproved background results the paper relies on.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Immediate 3D Gaussian Splat Reconstruction of Unordered Input with Global Consistency." pith.science (2026). https://pith.science/paper/PPHWO6AP

@misc{pith2026260714481,
  author       = {Pith},
  title        = {Pith review of: Immediate 3D Gaussian Splat Reconstruction of Unordered Input with Global Consistency},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PPHWO6AP}},
  note         = {Machine review of arXiv:2607.14481}
}
read the original abstract

3D Gaussian Splatting (3DGS) has become the method of choice for reconstructing and real-time rendering of captured scenes. To capture a scene with good visual quality, continuous image sequences are usually combined with out-of-order shots for better scene coverage. Structure from motion can reconstruct such captures, but only after they are all available and often with high computational cost. Incremental reconstruction methods -- often derived from SLAM solutions -- provide immediate feedback, but cannot handle the out-of-order capture we require. We provide the first immediate feedback solution for such radiance field capture that provides global consistency. We first introduce a method for fast matching in out-of-order sequences, by repurposing visual place recognition models and a covisibility graph, and provide an efficient way to find highly connected keyframes, improving quality even for ordered sequences. We show how these steps -- together with GPU optimization and careful Gaussian primitive placement -- provide fast local reconstruction, in our challenging radiance field reconstruction case. We then introduce a novel cluster-based method, again using the covisibility graph, to provide efficient loop closure that does not require sequential input. Finally, to handle large scenes in our context, we introduce a progressive hierarchy that allows our method to scale to large environments, without compromising efficiency. Our results show we provide immediate feedback 3DGS reconstruction with good visual quality in several datasets, with up to thousands of input images.

Figures

Figures reproduced from arXiv: 2607.14481 by the authors.

Figure 1
Figure 1. Our method takes an unordered sequence of input RGB images and directly computes a radiance field from them. To achieve this, we introduce a fast online pose estimation approach building on fast visual place recognition, local bundle adjustment, loop detection and loop closure techniques. Furthermore, a hierarchical Gaussian model is jointly optimized, allowing immediate feedback of reconstruction results during arb… view at source ↗
Figure 2
Figure 2. We show an intermediate sequence of registered frames for the [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Finding well-connected subsets with a greedy algorithm: From a [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Left we show cluster-based loop closure in the train scene, in a [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Loop Closure: (a) When a new keyframe (magenta) is matched [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Comparisons of our method with other pose-free solutions (see Sec. 8). We see that overall our method achieves higher visual quality, with sharper [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Comparisons of our method on the large scenes. Note that S3PO-GS and Octree-GS require a significantly larger amount of time for offline pose [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Overview of our proposed pipeline. Given an input image [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Loop Closure and Gaussian movement: (a) Through accumulated [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Additional comparisons. Fine-tuning [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: Scene with textureless areas and artificial repetitions [PITH_FULL_IMAGE:figures/full_fig_p016_11.png]
Figure 12
Figure 12. Figure 12: Limitations. (a) Without parallax (pure rotations), accurate triangulation is impossible, leading to poor geometry, visible in the depth renders (bottom [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

153 extracted references · 1 canonical work pages

  1. [1]

    ICCV , year=

    Longsplat: Robust unposed 3d gaussian splatting for casual long videos , author=. ICCV , year=

  2. [2]

    ACM Transactions on Graphics , year=

    Anysplat: Feed-forward 3d gaussian splatting from unconstrained views , author=. ACM Transactions on Graphics , year=

  3. [3]

    ICCV , year=

    Outdoor monocular slam with global scale-consistent 3d gaussian pointmaps , author=. ICCV , year=

  4. [4]

    Hughes and Y

    N. Hughes and Y. Chang and L. Carlone , booktitle=. Hydra: A Real-time Spatial Perception System for

  5. [5]

    Robotics , year=

    A comprehensive survey of visual slam algorithms , author=. Robotics , year=

  6. [6]

    , author=

    A survey of structure from motion*. , author=. Acta Numerica , year=

  7. [7]

    IPSJ transactions on computer vision and applications , year=

    Visual SLAM algorithms: A survey from 2010 to 2016 , author=. IPSJ transactions on computer vision and applications , year=

    2010

  8. [8]

    IEEE Transactions on Intelligent Transportation Systems , year=

    The revisiting problem in simultaneous localization and mapping: A survey on visual loop closure detection , author=. IEEE Transactions on Intelligent Transportation Systems , year=

Show all 153 references
  1. [9]

    Computer graphics forum , year=

    State of the art on neural rendering , author=. Computer graphics forum , year=

  2. [10]

    IEEE Transactions on Robotics , year=

    FrameSLAM: From bundle adjustment to real-time visual mapping , author=. IEEE Transactions on Robotics , year=

  3. [11]

    IEEE Robotics and Automation Letters , year=

    Bavle, Hriday and Sanchez-Lopez, Jose Luis and Shaheer, Muhammad and Civera, Javier and Voos, Holger , title=. IEEE Robotics and Automation Letters , year=

  4. [12]

    Vision, Modeling, and Visualization , title =

    Fink, Laura and Franke, Linus and Egger, Bernhard and Keinert, Joachim and Stamminger, Marc , year =. Vision, Modeling, and Visualization , title =

  5. [13]

    arXiv preprint arXiv:2511.10647 , year=

    Depth Anything 3: Recovering the visual space from any views , author=. arXiv preprint arXiv:2511.10647 , year=

    arXiv

  6. [14]

    ICCV , year=

    Generative novel view synthesis with 3d-aware diffusion models , author=. ICCV , year=

  7. [15]

    2023 , month =

    Agarwal, Sameer and Mierle, Keir and The Ceres Solver Team , title =. 2023 , month =

    2023

  8. [16]

    European Conference on Computer Vision (ECCV) , year=

    MegBA: A GPU-Based Distributed Library for Large-Scale Bundle Adjustment , author=. European Conference on Computer Vision (ECCV) , year=

  9. [17]

    IEEE International Conference on Robotics and Automation (ICRA) , year =

    g2o: A General Framework for Graph Optimization , author =. IEEE International Conference on Robotics and Automation (ICRA) , year =

  10. [18]

    , author=

    evo: Python package for the evaluation of odometry and SLAM. , author=

  11. [19]

    CVPR , year =

    Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields , author =. CVPR , year =

  12. [20]

    ICCV , year =

    Zip-NeRF: Anti-Aliased Grid-Based Neural Radiance Fields , author =. ICCV , year =

  13. [21]

    ACM Transactions on Graphics , year =

    Lombardi, Stephen and Simon, Tomas and Saragih, Jason and Schwartz, Gabriel and Lehrmann, Andreas and Sheikh, Yaser , title =. ACM Transactions on Graphics , year =

  14. [22]

    2024 , booktitle =

    Rip-NeRF: Anti-aliasing Radiance Fields with Ripmap-Encoded Platonic Solids , author =. 2024 , booktitle =

    2024

  15. [23]

    ACM Transactions on Graphics , year =

    Thomas M\"uller and Alex Evans and Christoph Schied and Alexander Keller , title =. ACM Transactions on Graphics , year =

  16. [24]

    European Conference on Computer Vision (ECCV) , year =

    Anpei Chen and Zexiang Xu and Andreas Geiger and Jingyi Yu and Hao Su , title =. European Conference on Computer Vision (ECCV) , year =

  17. [25]

    ICCV , year =

    Hu, Wenbo and Wang, Yuling and Ma, Lin and Yang, Bangbang and Gao, Lin and Liu, Xiao and Ma, Yuewen , title =. ICCV , year =

  18. [26]

    CVPR , year =

    Yu, Zehao and Chen, Anpei and Huang, Binbin and Sattler, Torsten and Geiger, Andreas , title =. CVPR , year =

  19. [27]

    Ecology , volume =

    Distance to nearest neighbor as a measure of spatial relationships in populations , author =. Ecology , volume =

  20. [28]

    2024 , eprint=

    CaRtGS: Computational Alignment for Real-Time Gaussian Splatting SLAM , author=. 2024 , eprint=

    2024

  21. [29]

    2024 , eprint=

    HI-SLAM2: Geometry-Aware Gaussian SLAM for Fast Monocular Scene Reconstruction , author=. 2024 , eprint=

    2024

  22. [30]

    IEEE Transactions on Visualization and Computer Graphics , year=

    3d gaussian splatting as new era: A survey , author=. IEEE Transactions on Visualization and Computer Graphics , year=

  23. [31]

    2024 , eprint=

    A Survey on 3D Gaussian Splatting , author=. 2024 , eprint=

    2024

  24. [32]

    arXiv preprint arXiv:2402.13255 , volume=

    How nerfs and 3d gaussian splatting are reshaping slam: a survey , author=. arXiv preprint arXiv:2402.13255 , volume=

    arXiv

  25. [33]

    2024 , eprint=

    ZeroGS: Training 3D Gaussian Splatting from Unposed Images , author=. 2024 , eprint=

    2024

  26. [34]

    ACM Transactions on Graphics , volume=

    Deep blending for free-viewpoint image-based rendering , author=. ACM Transactions on Graphics , volume=. 2018 , publisher=

    2018

  27. [35]

    Advances in Neural Information Processing Systems 32 , year =

    PyTorch: An Imperative Style, High-Performance Deep Learning Library , author =. Advances in Neural Information Processing Systems 32 , year =

  28. [36]

    Advances in neural information processing systems , volume =

    Hierarchical clustering of a mixture model , author =. Advances in neural information processing systems , volume =

  29. [37]

    Computer Graphics Forum , year =

    Progressive expectation-maximization for hierarchical volumetric photon mapping , author =. Computer Graphics Forum , year =

  30. [38]

    Wenbo, Chen and Ligang, Liu , title =. Proc. ACM Comput. Graph. Interact. Tech. (Proceedings of I3D 2024) , year =

    2024

  31. [39]

    and Tagliasacchi, Andrea , journal =

    Sabour, Sara and Goli, Lily and Kopanas, George and Matthews, Mark and Lagun, Dmitry and Guibas, Leonidas and Jacobson, Alec and Fleet, David J. and Tagliasacchi, Andrea , journal =

  32. [40]

    2025 , booktitle =

    Yang, Xijie and Xu, Linning and Jiang, Lihan and Lin, Dahua and Dai, Bo , title =. 2025 , booktitle =

    2025

  33. [41]

    European Conference on Computer Vision (ECCV) , year=

    CityGaussian: Real-time High-quality Large-Scale Scene Rendering with Gaussians , author=. European Conference on Computer Vision (ECCV) , year=

  34. [42]

    Global Structure-from-Motion Revisited , booktitle=

    Pan, Linfei and Baráth, Dániel and Pollefeys, Marc and Sch\". Global Structure-from-Motion Revisited , booktitle=

  35. [43]

    2024 , eprint=

    Octree-GS: Towards Consistent Real-time Rendering with LOD-Structured 3D Gaussians , author=. 2024 , eprint=

    2024

  36. [44]

    arXiv , year =

    Block-NeRF: Scalable Large Scene Neural View Synthesis , author =. arXiv , year =

  37. [45]

    ACM Transactions on Graphics , year =

    SMERF: Streamable Memory Efficient Radiance Fields for Real-Time Large-Scene Exploration , author =. ACM Transactions on Graphics , year =

  38. [46]

    ACM Transactions on Graphics , year =

    Kerbl, Bernhard and Meuleman, Andreas and Kopanas, Georgios and Wimmer, Michael and Lanvin, Alexandre and Drettakis, George , title =. ACM Transactions on Graphics , year =

  39. [47]

    CVPR , year =

    Turki, Haithem and Ramanan, Deva and Satyanarayanan, Mahadev , title =. CVPR , year =

  40. [48]

    CVPR , year =

    VastGaussian: Vast 3D Gaussians for Large Scene Reconstruction , author =. CVPR , year =

  41. [49]

    2023 , booktitle =

    Grid-guided Neural Radiance Fields for Large Urban Scenes , author =. 2023 , booktitle =

    2023

  42. [50]

    ICLR , year =

    Switch-NeRF: Learning Scene Decomposition with Mixture of Experts for Large-scale Neural Radiance Fields , author =. ICLR , year =

  43. [51]

    2024 , booktitle =

    CityGaussian: Real-time High-quality Large-Scale Scene Rendering with Gaussians , author =. 2024 , booktitle =

    2024

  44. [52]

    CVPR , year =

    Scaffold-gs: Structured 3d gaussians for view-adaptive rendering , author =. CVPR , year =

  45. [53]

    ICCV , year =

    Gao, Chen and Saraf, Ayush and Kopf, Johannes and Huang, Jia-Bin , title =. ICCV , year =

  46. [54]

    CVPR , month =

    Deng, Kangle and Liu, Andrew and Zhu, Jun-Yan and Ramanan, Deva , title =. CVPR , month =

  47. [55]

    2021 , booktitle =

    Robust Consistent Video Depth Estimation , author =. 2021 , booktitle =

    2021

  48. [56]

    European Conference on Computer Vision (ECCV) , year=

    Hac: Hash-grid assisted context for 3d gaussian splatting compression , author=. European Conference on Computer Vision (ECCV) , year=

  49. [57]

    2024 , isbn =

    Fang, Guangchi and Wang, Bing , title =. 2024 , isbn =. doi:10.1007/978-3-031-72980-5_10 , booktitle =

    2024 doi

  50. [58]

    arXiv preprint arXiv:2411.12788 , year=

    Mini-Splatting2: Building 360 Scenes within Minutes via Aggressive Gaussian Densification , author=. arXiv preprint arXiv:2411.12788 , year=

  51. [59]

    Papantonakis, Panagiotis and Kopanas, Georgios and Kerbl, Bernhard and Lanvin, Alexandre and Drettakis, George , booktitle =

  52. [60]

    2024 , booktitle =

    Mallick, Saswat Subhajyoti and Goel, Rahul and Kerbl, Bernhard and Steinberger, Markus and Carrasco, Francisco Vicente and De La Torre, Fernando , title =. 2024 , booktitle =

    2024

  53. [61]

    3D Gaussian Splatting for Real-Time Radiance Field Rendering , journal =

    Kerbl, Bernhard and Kopanas, Georgios and Leimk. 3D Gaussian Splatting for Real-Time Radiance Field Rendering , journal =

  54. [62]

    2020 , booktitle =

    NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis , author =. 2020 , booktitle =

    2020

  55. [63]

    ICCV , year =

    Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields , author =. ICCV , year =

  56. [64]

    2022 , booktitle =

    Plenoxels: Radiance Fields without Neural Networks , author =. 2022 , booktitle =

    2022

  57. [65]

    2020 , eprint =

    NeRF++: Analyzing and Improving Neural Radiance Fields , author =. 2020 , eprint =

    2020

  58. [66]

    ICCV , year=

    Outdoor Monocular SLAM with Global Scale-Consistent 3D Gaussian Pointmaps , author=. ICCV , year=

  59. [67]

    ACM Transactions on Graphics , year=

    On-the-fly Reconstruction for Large-Scale Novel View Synthesis from Unposed Images , author=. ACM Transactions on Graphics , year=

  60. [68]

    , booktitle=

    Hu, Yan Song and Abboud, Nicolas and Ali, Muhammad Qasim and Yang, Adam Srebrnjak and Elhajj, Imad and Asmar, Daniel and Chen, Yuhao and Zelek, John S. , booktitle=. MGSO: Monocular Real-Time Photometric SLAM with Efficient 3D Gaussian Splatting , year=

  61. [69]

    2025 , eprint=

    On-the-fly Large-scale 3D Reconstruction from Multi-Camera Rigs , author=. 2025 , eprint=

    2025

  62. [70]

    CVPR , year =

    Liu, Yu-Lun and Gao, Chen and Meuleman, Andreas and Tseng, Hung-Yu and Saraf, Ayush and Kim, Changil and Chuang, Yung-Yu and Kopf, Johannes and Huang, Jia-Bin , title =. CVPR , year =

  63. [71]

    and Kopf, Johannes , title =

    Meuleman, Andreas and Liu, Yu-Lun and Gao, Chen and Huang, Jia-Bin and Kim, Changil and Kim, Min H. and Kopf, Johannes , title =. CVPR , year =

  64. [72]

    ICCV , year =

    BARF: Bundle-Adjusting Neural Radiance Fields , author =. ICCV , year =

  65. [73]

    ICCV , year =

    Self-Calibrating Neural Radiance Fields , author =. ICCV , year =

  66. [74]

    2026 , eprint=

    VPGS-SLAM: Voxel-based Progressive 3D Gaussian SLAM in Large-Scale Scenes , author=. 2026 , eprint=

    2026

  67. [75]

    IEEE International Conference on Robotics and Automation (ICRA) , year=

    Large-Scale Gaussian Splatting SLAM , author=. IEEE International Conference on Robotics and Automation (ICRA) , year=

  68. [76]

    TLS-SLAM: Gaussian Splatting SLAM Tailored for Large-Scale Scenes , year=

    Cheng, Sicong and He, Songyang and Duan, Fuqing and An, Ning , journal=. TLS-SLAM: Gaussian Splatting SLAM Tailored for Large-Scale Scenes , year=

  69. [77]

    2024 , eprint =

    InstantSplat: Unbounded Sparse-view Pose-free Gaussian Splatting in 40 Seconds , author =. 2024 , eprint =

    2024

  70. [78]

    2024 , booktitle =

    GGRt: Towards Generalizable 3D Gaussians without Pose Priors in Real-Time , author =. 2024 , booktitle =

    2024

  71. [79]

    ACM SIGGRAPH Conference Proceedings , year =

    A Construct-Optimize Approach to Sparse View Synthesis without Camera Pose , author =. ACM SIGGRAPH Conference Proceedings , year =

  72. [80]

    Hidenobu Matsuki and Riku Murai and Paul H. J. Kelly and Andrew J. Davison , booktitle =

  73. [81]

    CVPR , year =

    Photo-SLAM: Real-time Simultaneous Localization and Photorealistic Mapping for Monocular, Stereo, and RGB-D Cameras , author =. CVPR , year =

  74. [82]

    and Wang, Xiaolong , title =

    Fu, Yang and Liu, Sifei and Kulkarni, Amey and Kautz, Jan and Efros, Alexei A. and Wang, Xiaolong , title =. CVPR , year =

  75. [83]

    3DV , year =

    Liyuan Zhu and Yue Li and Erik Sandström and Shengyu Huang and Konrad Schindler and Iro Armeni , title =. 3DV , year =

  76. [84]

    CVPR , year =

    SplaTAM: Splat, Track & Map 3D Gaussians for Dense RGB-D SLAM , author =. CVPR , year =

  77. [85]

    ACM SIGGRAPH Conference Proceedings , year =

    Zhexi Peng and Tianjia Shao and Liu Yong and Jingke Zhou and Yin Yang and Jingdong Wang and Kun Zhou , title =. ACM SIGGRAPH Conference Proceedings , year =

  78. [86]

    Lilienthal and Martin Magnusson , title =

    Shuo Sun and Malcolm Mielle and Achim J. Lilienthal and Martin Magnusson , title =. CVPR , year =

  79. [87]

    Lilienthal and Martin Magnusson , title =

    Shuo Sun and Malcolm Mielle and Achim J. Lilienthal and Martin Magnusson , title =. IROS , year =

  80. [88]

    2024 , eprint =

    Compact 3D Gaussian Splatting For Dense Visual SLAM , author =. 2024 , eprint =

    2024

  81. [89]

    and Al-Hamadi, Ayoub , booktitle=

    Hempel, Thorsten and Abdelrahman, Ahmed A. and Al-Hamadi, Ayoub , booktitle=. 6d Rotation Representation For Unconstrained Head Pose Estimation , year=

  82. [90]

    CVPR , year=

    On the Continuity of Rotation Representations in Neural Networks , author=. CVPR , year=

  83. [91]

    ORB: An efficient alternative to SIFT or SURF , year =

    Rublee, Ethan and Rabaud, Vincent and Konolige, Kurt and Bradski, Gary , booktitle =. ORB: An efficient alternative to SIFT or SURF , year =

  84. [92]

    European Conference on Computer Vision (ECCV) , year =

    Multiview Stereo with Cascaded Epipolar RAFT , author =. European Conference on Computer Vision (ECCV) , year =

  85. [93]

    CVPR , title =

    Guilherme. CVPR , title =

  86. [94]

    ICCV , year =

    Philipp Lindenberger and Paul-Edouard Sarlin and Marc Pollefeys , title =. ICCV , year =

  87. [95]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

    Ali-bey, Amar and Chaib-draa, Brahim and Gigu. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

  88. [96]

    CVPR , year =

    Barath, Daniel and Noskova, Jana and Ivashechkin, Maksym and Matas, Jiri , title =. CVPR , year =

  89. [97]

    DAGM-Symposium , year =

    Locally Optimized RANSAC , author =. DAGM-Symposium , year =

  90. [98]

    2009 , booktitle =

    EPnP: An Accurate O(n) Solution to the PnP Problem , author =. 2009 , booktitle =

    2009

  91. [99]

    European Conference on Computer Vision (ECCV) , month =

    Persson, Mikael and Nordberg, Klas , title =. European Conference on Computer Vision (ECCV) , month =

  92. [100]

    , journal =

    Nister, D. , journal =. An efficient solution to the five-point relative pose problem , year =

  93. [101]

    1991 , publisher=

    Computer and robot vision , author=. 1991 , publisher=

    1991

  94. [102]

    Hartley, R. I. and Zisserman, A. , title =. 2004 , publisher =

    2004

  95. [103]

    Quarterly of Applied Mathematics , year =

    A Method for the Solution of Certain Non-Linear Problems in Least Squares , author =. Quarterly of Applied Mathematics , year =

  96. [104]

    Marquardt , journal =

    Donald W. Marquardt , journal =. An Algorithm for Least-Squares Estimation of Nonlinear Parameters , year =

  97. [105]

    Madsen, Kaj and Nielsen, Hans and Tingleff, O , year =

  98. [106]

    Jeon and Min H

    Andreas Meuleman and Hyeonjoong Jang and Daniel S. Jeon and Min H. Kim , title =. CVPR , year =

  99. [107]

    International Conference on 3D Vision (3DV) , year=

    3d reconstruction with spatial memory , author=. International Conference on 3D Vision (3DV) , year=

  100. [108]

    Advances in Neural Information Processing Systems , year=

    VGGT-SLAM: Dense RGB SLAM Optimized on the SL (4) Manifold , author=. Advances in Neural Information Processing Systems , year=

  101. [109]

    , booktitle=

    Murai, Riku and Dexheimer, Eric and Davison, Andrew J. , booktitle=

  102. [110]

    , journal=

    Galvez-López, Dorian and Tardos, Juan D. , journal=. Bags of Binary Words for Fast Place Recognition in Image Sequences , year=

  103. [111]

    Mur-Artal, Ra\'ul and Montiel, J. M. M. and Tard\'os, Juan D. , journal=

  104. [112]

    , journal=

    Mur-Artal, Ra\'ul and Tard\'os, Juan D. , journal=

  105. [113]

    and Montiel, Jos\'e M

    Campos, Carlos and Elvira, Richard and G\`omez, Juan J. and Montiel, Jos\'e M. M. and Tard\'os, Juan D. , journal =

  106. [114]

    and Montiel, J.M.M

    Strasdat, Hauke and Davison, Andrew J. and Montiel, J.M.M. and Konolige, Kurt , booktitle=. Double window optimisation for constant time visual SLAM , year=

  107. [115]

    Robotics: Science and Systems , year=

    Scale Drift-Aware Large Scale Monocular SLAM , author=. Robotics: Science and Systems , year=

  108. [116]

    Wang, Chen and Gao, Dasong and Xu, Kuan and Geng, Junyi and Hu, Yaoyu and Qiu, Yuheng and Li, Bowen and Yang, Fan and Moon, Brady and Pandey, Abhinav and Aryan and Xu, Jiahe and Wu, Tianhao and He, Haonan and Huang, Daning and Ren, Zhongqiang and Zhao, Shibo and Fu, Taimeng an...

  109. [117]

    International Journal of Robotics Research , year=

    ElasticFusion: Real-Time Dense SLAM and Light Source Estimation , author=. International Journal of Robotics Research , year=

  110. [118]

    , title =

    Liso, Lorenzo and Sandstr\"om, Erik and Yugay, Vladimir and Van Gool, Luc and Oswald, Martin R. , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2024 , pages =

    2024

  111. [119]

    Proceedings of the SIGGRAPH Asia 2025 Conference Papers , pages=

    EGG-Fusion: Efficient 3D Reconstruction with Geometry-aware Gaussian Surfel on the Fly , author=. Proceedings of the SIGGRAPH Asia 2025 Conference Papers , pages=

    2025

  112. [120]

    Proceedings of the SIGGRAPH Asia 2025 Conference Papers , pages=

    Gigaslam: Large-scale monocular slam with hierarchical gaussian splats , author=. Proceedings of the SIGGRAPH Asia 2025 Conference Papers , pages=

    2025

  113. [121]

    Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers , pages=

    Monocular online reconstruction with enhanced detail preservation , author=. Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers , pages=

  114. [122]

    Robotics: Science and Systems , year=

    ElasticFusion: Dense SLAM Without A Pose Graph , author=. Robotics: Science and Systems , year=

  115. [123]

    CVPR , year =

    Schops, Thomas and Sattler, Torsten and Pollefeys, Marc , title =. CVPR , year =

  116. [124]

    IEEE Transactions on Robotics , year =

    Mur-Artal, Ra. IEEE Transactions on Robotics , year =

  117. [125]

    arXiv e-prints , pages=

    DROID-Splat: Combining end-to-end SLAM with 3D Gaussian Splatting , author=. arXiv e-prints , pages=

  118. [126]

    NIPS , year =

    Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras , author =. NIPS , year =

  119. [127]

    ICLR , year =

    DeepV2D: Video to Depth with Differentiable Structure from Motion , author =. ICLR , year =

  120. [128]

    DeepFactors: Real-time probabilistic dense monocular SLAM , year =

    Czarnowski, J and Laidlow, T and Clark, R and Davison, AJ , journal =. DeepFactors: Real-time probabilistic dense monocular SLAM , year =

  121. [129]

    CVPR , year=

    VGGT: Visual Geometry Grounded Transformer , author=. CVPR , year=

  122. [130]

    Wang, Yifan and Zhou, Jianjun and Zhu, Haoyi and Chang, Wenzheng and Zhou, Yang and Li, Zizun and Chen, Junyi and Pang, Jiangmiao and Shen, Chunhua and He, Tong , journal=. ^

  123. [131]

    Walter and Vitor Campagnolo Guizilini and Greg Shakhnarovich , title =

    Jiahao Li and Haochen Wang and Muhammad Zubair Irshad and Igor Vasiljevic and Matthew R. Walter and Vitor Campagnolo Guizilini and Greg Shakhnarovich , title =. 3DV , year =

  124. [132]

    European Conference on Computer Vision (ECCV) , year=

    Scene Coordinate Reconstruction: Posing of Image Collections via Incremental Learning of a Relocalizer , author=. European Conference on Computer Vision (ECCV) , year=

  125. [133]

    Robust consistent video depth estimation , author =

  126. [134]

    European Conference on Computer Vision (ECCV) , year =

    ParticleSfM: Exploiting Dense Point Trajectories for Localizing Moving Cameras in the Wild , author =. European Conference on Computer Vision (ECCV) , year =

  127. [135]

    Advances in Neural Information Processing Systems , year=

    Depth Anything V2 , author=. Advances in Neural Information Processing Systems , year=

  128. [136]

    ACM siggraph 2006 papers , pages=

    Photo tourism: exploring photo collections in 3D , author=. ACM siggraph 2006 papers , pages=

    2006

  129. [137]

    Structure-from-Motion Revisited , booktitle =

    Sch\". Structure-from-Motion Revisited , booktitle =

  130. [138]

    CVPR , year =

    Wang, Shuzhe and Leroy, Vincent and Cabon, Yohann and Chidlovskii, Boris and Revaud, Jerome , title =. CVPR , year =

  131. [139]

    2024 , eprint =

    MASt3R-SfM: a Fully-Integrated Solution for Unconstrained Structure-from-Motion , author =. 2024 , eprint =

    2024

  132. [140]

    arXiv preprint arXiv:2408.16061 , year =

    3D Reconstruction with Spatial Memory , author =. arXiv preprint arXiv:2408.16061 , year =

    arXiv

  133. [141]

    Pixelwise View Selection for Unstructured Multi-View Stereo , booktitle =

    Sch\". Pixelwise View Selection for Unstructured Multi-View Stereo , booktitle =

  134. [142]

    European Conference on Computer Vision (ECCV) , year =

    LSD-SLAM: Large-scale direct monocular SLAM , author =. European Conference on Computer Vision (ECCV) , year =

  135. [143]

    Direct sparse odometry , author =

  136. [144]

    IEEE transactions on robotics , year =

    ORB-SLAM: a versatile and accurate monocular SLAM system , author =. IEEE transactions on robotics , year =

  137. [145]

    IEEE transactions on robotics , year =

    Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras , author =. IEEE transactions on robotics , year =

  138. [146]

    DTAM: Dense tracking and mapping in real-time , author =

  139. [147]

    Unsupervised learning of depth and ego-motion from video , author =

  140. [148]

    Digging into self-supervised monocular depth estimation , author =

  141. [149]

    CVPR , year=

    A benchmark dataset and evaluation methodology for video object segmentation , author=. CVPR , year=

  142. [150]

    A benchmark for the evaluation of RGB-D SLAM systems , year =

    Sturm, Jürgen and Engelhard, Nikolas and Endres, Felix and Burgard, Wolfram and Cremers, Daniel , booktitle =. A benchmark for the evaluation of RGB-D SLAM systems , year =

  143. [151]

    The EuRoC micro aerial vehicle datasets , journal =

    Burri, Michael and Nikolic, Janosch and Gohl, Pascal and Schneider, Thomas and Rehder, Joern and Omari, Sammy and Achtelik, Markus and Siegwart, Roland , year =. The EuRoC micro aerial vehicle datasets , journal =

  144. [152]

    Julian Straub and Thomas Whelan and Lingni Ma and Yufan Chen and Erik Wijmans and Simon Green and Jakob J. Engel and Raul Mur-Artal and Carl Ren and Shobhit Verma and Anton Clarkson and Mingfei Yan and Brian Budge and Yajie Yan and Xiaqing Pan and June Yon and Yuyang Zou and K...

  145. [153]

    2017 , journal =

    Knapitsch, Arno and Park, Jaesik and Zhou, Qian-Yi and Koltun, Vladlen , title =. 2017 , journal =

    2017

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.