Pith. sign in

REVIEW 2 major objections 5 minor 75 references

Glob3R turns coarse feed-forward 3D foundation predictions into multi-view tracks and refines them with global motion averaging and bundle adjustment for accurate, scalable reconstruction.

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 ↗

A frozen Pi3X backbone plus dense warping tracks and keyframe sliding-window global optimization yields more accurate, scalable SfM than feed-forward or classical baselines alone.

T0 review reviewed 2026-07-13 challenge →

load-bearing objection Solid hybrid systems paper: frozen Pi3X priors + dense-warp tracks + global BA beat feed-forward and many streaming baselines, but “consistent” gains partly rest on scene-tuned window size. the 2 major comments →

arxiv 2607.09225 v1 pith:L5TPKPSJ submitted 2026-07-10 cs.CV

Glob3R: Global Structure-from-Motion with 3D Foundation Models

Junyuan Deng , Heng Li , Kejie Qiu , Lingteng Qiu , Rui Peng , Weichao Shen , Weihao Yuan , Siyu Zhu
show 2 more authors
Zilong Dong Ping Tan
This is my paper
classification cs.CV
keywords structure-from-motion3D foundation modelsdense matchingfeature tracksmotion averagingbundle adjustmentsliding-window associationneural rendering
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 reading

Feed-forward 3D foundation models can guess camera poses and dense geometry from image sets, but the guesses are often only approximate and hard to scale without drift when scenes are long or unordered. This paper argues that those predictions become much more useful once they are converted into explicit multi-view correspondences and optimized like classical global structure-from-motion. A frozen foundation backbone is given a light dense-matching head that warps keyframes to neighboring views; the warps yield sparse reliable feature tracks, which a keyframe sliding-window scheme propagates across overlapping windows. Global rotation and translation averaging, then bundle adjustment, resolve scale inconsistencies and refine every camera. On indoor, outdoor, driving, and unordered benchmarks the refined poses are more accurate and more stable than pure feed-forward baselines, recent streaming methods, and classical SfM, and they improve neural rendering quality.

Core claim

The central claim is that foundation-model geometric priors become high-accuracy, globally consistent reconstructions when dense warps are turned into multi-view tracks and optimized with classical global SfM (motion averaging plus bundle adjustment), rather than by stitching independent chunks. This hybrid is more accurate than feed-forward foundation models and recent scalable learners, and more robust than classical SfM on the reported benchmarks, including higher novel-view synthesis quality from the refined poses.

What carries the argument

Keyframe-based sliding-window dense warping on a frozen foundation backbone: dense image warps from selected keyframes produce sparse multi-view tracks that seed a pose graph; overlapping windows propagate tracks and relative poses, then rotation/translation averaging and bundle adjustment refine poses, scales, and dense geometry at frame level.

Load-bearing premise

The frozen foundation model’s local geometry and features must already be good enough that a light matching head can extract reliable tracks; when the initial geometry is badly ambiguous, matches go wrong and optimization cannot fully recover.

What would settle it

On a long sequence or unordered set where the foundation model’s initial point maps and poses are deliberately degraded or known to be fragmented (as in the paper’s Ballroom-style failure), measure whether track-based global averaging and bundle adjustment still recover low trajectory error and high relative-pose accuracy; if error remains near the corrupted initialization, the hybrid claim fails.

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

If this is right

  • Feed-forward foundation reconstructions can be upgraded for high-fidelity uses such as neural rendering without retraining the backbone.
  • Long ordered sequences and large unordered collections can be handled by overlapping windows and track propagation instead of independent chunk stitching.
  • Pose graphs built from foundation-guided dense warps are more stable for global averaging than graphs from classical matching alone in hard texture or forward-motion settings.
  • Recovered sparse tracks plus rescaled dense depths yield globally aligned dense point clouds after optimization.

Where Pith is reading between the lines

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

  • Adaptive window size driven by overlap and matching confidence, which the paper flags as future work, would likely reduce the remaining failures on scenes where fixed windows give the backbone too little context.
  • If matching can be made less dependent on imperfect backbone geometry, the same pipeline could serve as a general post-optimizer for other feed-forward 3D models beyond the one used here.
  • The gains on novel-view synthesis suggest that many NeRF/view-synthesis failures currently blamed on the renderer may instead be fixed by this style of pose refinement.
Share X Bluesky LinkedIn Reddit HN

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 / 5 minor

Summary. Glob3R converts feed-forward predictions of a frozen Pi3X 3D foundation model into optimizable multi-view constraints for global Structure-from-Motion. A lightweight dense matching head predicts image warps from keyframes to neighboring views; these warps are sparsified into multi-view tracks. A keyframe-based sliding-window association strategy with half-window overlap builds a pose graph over long sequences and unordered collections (via retrieval). Global rotation/translation averaging followed by bundle adjustment then refines poses, reduces scale drift, and recovers dense geometry. Experiments on Tanks and Temples (novel-view PSNR via Nerfacto), TUM RGB-D and KITTI (trajectory RMSE), and ETH3D (RRA/RTA) show consistent gains over feed-forward foundation models, recent streaming/chunk methods, and classical SfM, with ablations isolating motion averaging, full BA, and the matching head.

Significance. If the reported gains hold under fixed inference settings, the work is a clear and useful contribution: it shows how to turn foundation-model geometric priors and intermediate tokens into reliable tracks for classical global SfM, rather than merely stitching chunk-level SE(3)/Sim(3) alignments. Strengths include multi-benchmark evaluation spanning indoor, outdoor, driving, and unordered SfM settings; an ablation of initialization vs. motion averaging vs. full BA and of matching variants (coarse, VGGT-track, RoMaV2); and explicit failure analysis when backbone geometry is ambiguous. The combination of frozen foundation priors with frame-level global optimization is practically relevant for high-fidelity pose estimation and neural rendering, and the pipeline is more flexible than methods that retrain the backbone or assume strictly sequential input.

major comments (2)
  1. Appendix C.1 states that the default sliding window is N=20, but N is raised to 120 for T&T Auditorium and Courtroom and to 200 for KITTI sequence 02 “to stabilize Pi3X predictions.” Those scenes appear in the headline results (full T&T PSNR Table 6; KITTI Table 3). Because matching and tracks are built on frozen backbone tokens H and predicted point maps (§3.1–3.2, Eqs. 1–2), enlarging the window changes the multi-view context that produces the geometric prior the matching head depends on. Without a fixed-N ablation (or at least fixed-N numbers for the adjusted scenes), it is unclear how much of the claimed “consistent” improvement is the optimization pipeline versus scene-tuned context for the backbone. This is load-bearing for the central claim of consistent gains over feed-forward and streaming baselines.
  2. The pipeline’s dependence on Pi3X’s initial geometry is acknowledged in the Ballroom failure case (Appendix C.4 / Fig. 8): ambiguous foundation-model structure corrupts warps and BA cannot fully recover. The same premise underlies keyframe selection (Eq. 4) and track generation from predicted warps/point maps. The paper would be stronger if it quantified how often this failure mode occurs (e.g., fraction of scenes/windows with large residual scale or track outlier rates) and whether retrieval-based loop edges or denser tracks mitigate it, rather than treating it only as a qualitative limitation.
minor comments (5)
  1. T&T evaluation uses Nerfacto PSNR as a pose proxy because accurate GT poses are unavailable (§4.1). This is reasonable and follows prior work, but the paper should state more clearly that PSNR can also reflect depth/scale consistency and rendering hyperparameters, not pure pose error alone.
  2. Table 5 reports matching ablations only on ETH3D. A short note on whether the same ranking (dense warp > VGGT-track / RoMaV2) holds on sequential benchmarks would strengthen the matching-head claim.
  3. Notation: Eq. (1) uses m_i for metric scale while later text refers to “predicted metric scale”; keep symbols consistent. Also clarify whether relative poses on graph edges are taken from Pi3X T_i or re-estimated from tracks before motion averaging.
  4. Runtime Table 7 is useful; adding a breakdown (backbone forward, matching, association, motion averaging, BA) would help readers judge scalability of the optimization stage.
  5. Minor presentation: “Glob3R” vs “Glob3r” in the project URL; a few repeated COLMAP citations ([42]/[43]); and ensure all baseline configurations in Appendix C.2 are fully reproducible (e.g., exact SALAD threshold and track sample counts).

Circularity Check

0 steps flagged

No significant circularity: supervised matching and classical global SfM are evaluated on external benchmarks, not forced by construction.

full rationale

Glob3R’s chain is: freeze Pi3X geometric tokens/poses/point maps → train a dense warp head with GT 2D–2D warps from depth/poses (Eqs. 2–3, App. B) → convert warps to multi-view tracks via keyframe sliding windows → motion averaging + BA (Eqs. 5–6) → report pose/NVS metrics on TUM, KITTI, ETH3D, and T&T. Training supervision is standard multi-view learning; it does not define the test metrics. Reported gains are measured against external ground-truth trajectories, relative-pose thresholds, and novel-view PSNR, with ablations (Table 5) isolating matching and BA. Self-citations (e.g., SAIL-Recon) appear as baselines, not as uniqueness theorems or load-bearing premises that force the result. Scene-dependent window sizes (App. C.1) and backbone-geometry failure modes (App. C.4) are experimental-design and robustness issues, not circular reductions of predictions to fitted inputs. No self-definitional equations, fitted-input-as-prediction, or ansatz-via-self-citation steps appear in the derivation.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 2 invented entities

The central empirical claim rests on standard multi-view geometry plus several engineering choices: frozen Pi3X priors, a trained dense matching head, hand-set association thresholds, and classical robust averaging/BA. No new physical entities; free parameters are algorithmic thresholds and training hyperparameters that affect reported accuracy, especially window size on hard scenes.

free parameters (6)
  • sliding_window_size_N
    Default N=20 (stride 10); raised to 120 or 200 on specific hard sequences to stabilize Pi3X. Directly affects geometry quality and reported results.
  • keyframe_reprojection_threshold
    Frames become keyframes when valid projected pixel ratio falls below ~0.2 (appendix); controls track graph density.
  • warp_and_depth_confidence_thresholds
    Warp conf 0.6 and depth conf 0.1 filter tracks; change which correspondences enter averaging/BA.
  • tracks_per_keyframe
    Default 512 samples (256 on large KITTI); free sampling budget for sparse tracks.
  • matching_loss_weights_lambda
    λ_NLL, λ_warp, λ_conf balance coarse NLL, warp regression, and confidence BCE during matching-head training.
  • retrieval_similarity_threshold
    SALAD cosine similarity >0.5 for loop/long-range keyframe windows; affects global graph connectivity.
axioms (4)
  • standard math Multi-view geometry: relative poses and tracks can be refined by rotation/translation averaging and reprojection-based bundle adjustment under robust losses.
    Invoked throughout §3.3 (Eqs. 5–6) following GLOMAP-style global SfM.
  • domain assumption Frozen Pi3X intermediate features encode useful cross-view geometric cues for dense correspondence without unfreezing the backbone.
    Core design of §3.1; matching decoder operates on frozen H tokens.
  • domain assumption Half-overlapping sliding windows plus keyframe tracks suffice to propagate constraints across long sequences and retrieval-ordered unordered sets.
    §3.2 sequence/window construction and track merging into a pose graph.
  • ad hoc to paper Dense image warping (coarse DPT + RoMa-style refinement) yields more reliable multi-view tracks than VGGT point tracking or pure two-view RoMa for this pipeline.
    Architectural choice justified by ablation Table 5, not a general theorem.
invented entities (2)
  • Glob3R dense matching head on frozen Pi3X tokens no independent evidence
    purpose: Predict multi-view image warps and confidences to form sparse tracks for global optimization without retraining the foundation model.
    New module (Decmatch + DPTmatch + refinement) trained while freezing Pi3X; independent evidence is empirical matching/pose gains, not external physical prediction.
  • Keyframe-based sliding-window track association graph no independent evidence
    purpose: Scale reconstruction by propagating tracks and relative poses across overlapping windows and retrieval loops at frame level rather than chunk stitching.
    Pipeline construct in §3.2; validated only within this paper’s experiments.

reviewed 2026-07-13 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Glob3R: Global Structure-from-Motion with 3D Foundation Models." pith.science (2026). https://pith.science/paper/L5TPKPSJ

@misc{pith2026260709225,
  author       = {Pith},
  title        = {Pith review of: Glob3R: Global Structure-from-Motion with 3D Foundation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L5TPKPSJ}},
  note         = {Machine review of arXiv:2607.09225}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Recent 3D geometric foundation models, such as VGGT, provide robust feed-forward 3D reconstruction by directly predicting camera poses and 3D scene points from input images. However, their results remain inaccurate, and scaling them to long sequences or large unordered image sets typically requires chunk-wise processing, which can introduce drift and inconsistency. We present Glob3R, a global SfM-style reconstruction built on 3D foundation models. Our key idea is to explicitly optimize feed-forward geometric predictions. To this end, we augment a frozen Pi3X backbone with a lightweight dense matching head that predicts image warps between selected reference frames and neighboring views. These dense warps are converted into sparse but reliable multi-view feature tracks, which provide correspondence constraints for global optimization. We further introduce a keyframe-based sliding-window association strategy that propagates tracks and relative poses across overlapping windows, enabling scalable reconstruction. Finally, we perform global motion averaging and bundle adjustment to refine camera poses, reduce scale inconsistencies, and recover dense scene geometry. Extensive experiments on indoor, outdoor, large-scale driving, and unordered SfM benchmarks demonstrate that Glob3R achieves robust and accurate reconstruction. It consistently improves over feed-forward foundation-model baselines and recent scalable reconstruction methods, while being more robust than classical SfM pipelines. The refined poses also lead to higher-quality neural rendering, validating the benefit of combining foundation-model priors with global geometric optimization. Project page: https://junyuandeng.github.io/Glob3r

Figures

Figures reproduced from arXiv: 2607.09225 by Heng Li, Junyuan Deng, Kejie Qiu, Lingteng Qiu, Ping Tan, Rui Peng, Siyu Zhu, Weichao Shen, Weihao Yuan, Zilong Dong.

Figure 1
Figure 1. Figure 1: Overview of the proposed reconstruction framework. Given an ordered image sequence or a retrieval-based pseudo-sequence, we first predict geometric priors and dense warps within windows. Keyframes are selected according to reprojection coverage, and dense warps are converted into sparse multi-view tracks. These tracks are then merged into a global association graph for rotation averaging, translation avera… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative comparison of novel-view synthesis on Tanks and Temples. Our method [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Camera trajectory comparison on KITTI Odometry. Our method better preserves the global [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative results on various datasets. We visualize the original images, warped results, [PITH_FULL_IMAGE:figures/full_fig_p019_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Additional qualitative comparisons on Tanks and Temples. Our method produces sharper [PITH_FULL_IMAGE:figures/full_fig_p020_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Additional KITTI trajectory visualizations. Our method better preserves global trajectory [PITH_FULL_IMAGE:figures/full_fig_p021_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Dense point cloud reconstructions produced by our optimized poses and recovered dense [PITH_FULL_IMAGE:figures/full_fig_p022_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Failure case on the T&T Ballroom [24] scene. Ambiguous foundation-model geometry leads to incorrect matching and degraded reconstruction. Runtime Analyse [PITH_FULL_IMAGE:figures/full_fig_p025_8.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

75 extracted references · 19 linked inside Pith

  1. [1]

    Arkitscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data.arXiv preprint arXiv:2111.08897, 2021

    Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data.arXiv preprint arXiv:2111.08897, 2021

    Pith/arXiv arXiv 2021

  2. [2]

    Scene coordinate reconstruction: Posing of image collections via incremental learning of a relocalizer

    Eric Brachmann, Jamie Wynn, Shuai Chen, Tommaso Cavallari, Áron Monszpart, Daniyar Turmukham- betov, and Victor Adrian Prisacariu. Scene coordinate reconstruction: Posing of image collections via incremental learning of a relocalizer. InECCV, 2024

    2024

  3. [3]

    Virtual kitti 2.arXiv preprint arXiv:2001.10773, 2020

    Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual kitti 2.arXiv preprint arXiv:2001.10773, 2020

    Pith/arXiv arXiv 2001

  4. [4]

    Geometric context transformer for streaming 3d reconstruction.arXiv preprint arXiv:2604.14141, 2026

    Lin-Zhuo Chen, Jian Gao, Yihang Chen, Ka Leong Cheng, Yipengjing Sun, Liangxiao Hu, Nan Xue, Xing Zhu, Yujun Shen, Yao Yao, and Yinghao Xu. Geometric context transformer for streaming 3d reconstruction.arXiv preprint arXiv:2604.14141, 2026

    Pith/arXiv arXiv 2026

  5. [5]

    Longstream: Long-sequence streaming autoregressive visual geometry.arXiv preprint arXiv:2602.13172, 2026

    Chong Cheng, Xianda Chen, Tao Xie, Wei Yin, Weiqiang Ren, Qian Zhang, Xiaoyang Guo, and Hao Wang. Longstream: Long-sequence streaming autoregressive visual geometry.arXiv preprint arXiv:2602.13172, 2026

    arXiv 2026

  6. [6]

    Scannet: Richly-annotated 3d reconstructions of indoor scenes

    Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 5828–5839, 2017

    2017

  7. [7]

    Nerf-loam: Neural implicit representation for large-scale incremental lidar odometry and mapping

    Junyuan Deng, Qi Wu, Xieyuanli Chen, Songpengcheng Xia, Zhen Sun, Guoqing Liu, Wenxian Yu, and Ling Pei. Nerf-loam: Neural implicit representation for large-scale incremental lidar odometry and mapping. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 8218–8227, 2023

    2023

  8. [8]

    Boost 3d reconstruction using diffusion-based monocular camera calibration

    Junyuan Deng, Wei Yin, Xiaoyang Guo, Qian Zhang, Xiaotao Hu, Weiqiang Ren, Xiao-Xiao Long, and Ping Tan. Boost 3d reconstruction using diffusion-based monocular camera calibration. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 7110–7121, 2025

    2025

  9. [9]

    Sail-recon: Large sfm by augmenting scene regression with localization.2026 International Conference on 3D Vision (3DV)., 2026

    Junyuan Deng, Heng Li, Tao Xie, Weiqiang Ren, Qian Zhang, Ping Tan, and Xiaoyang Guo. Sail-recon: Large sfm by augmenting scene regression with localization.2026 International Conference on 3D Vision (3DV)., 2026

    2026

  10. [10]

    Vggt-long: Chunk it, loop it, align it – pushing vggt’s limits on kilometer-scale long rgb sequences, 2025

    Kai Deng, Zexin Ti, Jiawei Xu, Jian Yang, and Jin Xie. Vggt-long: Chunk it, loop it, align it – pushing vggt’s limits on kilometer-scale long rgb sequences, 2025. URL https://arxiv.org/abs/2507.16443

    Pith/arXiv arXiv 2025

  11. [11]

    Reloc-vggt: Visual re-localization with geometry grounded transformer.arXiv preprint arXiv:2512.21883, 2025

    Tianchen Deng, Wenhua Wu, Kunzhen Wu, Guangming Wang, Siting Zhu, Shenghai Yuan, Xun Chen, Guole Shen, Zhe Liu, and Hesheng Wang. Reloc-vggt: Visual re-localization with geometry grounded transformer.arXiv preprint arXiv:2512.21883, 2025. 10

    arXiv 2025

  12. [12]

    Mast3r-sfm: a fully-integrated solution for unconstrained structure-from-motion.arXiv preprint arXiv:2409.19152, 2024

    Bardienus Duisterhof, Lojze Zust, Philippe Weinzaepfel, Vincent Leroy, Yohann Cabon, and Jerome Revaud. Mast3r-sfm: a fully-integrated solution for unconstrained structure-from-motion.arXiv preprint arXiv:2409.19152, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    Dkm: Dense kernelized feature matching for geometry estimation

    Johan Edstedt, Ioannis Athanasiadis, Mårten Wadenbäck, and Michael Felsberg. Dkm: Dense kernelized feature matching for geometry estimation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17765–17775, 2023

    2023

  14. [14]

    Roma: Robust dense feature matching

    Johan Edstedt, Qiyu Sun, Georg Bökman, Mårten Wadenbäck, and Michael Felsberg. Roma: Robust dense feature matching. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19790–19800, 2024

    2024

  15. [15]

    Roma v2: Harder better faster denser feature matching.arXiv preprint arXiv:2511.15706, 2025

    Johan Edstedt, David Nordström, Yushan Zhang, Georg Bökman, Jonathan Astermark, Viktor Larsson, Anders Heyden, Fredrik Kahl, Mårten Wadenbäck, and Michael Felsberg. Roma v2: Harder better faster denser feature matching.arXiv preprint arXiv:2511.15706, 2025

    Pith/arXiv arXiv 2025

  16. [16]

    More: 3d visual geometry reconstruction meets mixture-of-experts

    Jingnan Gao, Zhe Wang, Xianze Fang, Xingyu Ren, Zhuo Chen, Shengqi Liu, Yuhao Cheng, Jiangjing Lyu, Xiaokang Yang, and Yichao Yan. More: 3d visual geometry reconstruction meets mixture-of-experts. arXiv preprint arXiv:2510.27234, 2025

    arXiv 2025

  17. [17]

    Are we ready for autonomous driving? The KITTI vision benchmark suite

    Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? The KITTI vision benchmark suite. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2012

    2012

  18. [18]

    Combining two-view constraints for motion estimation

    Venu Madhav Govindu. Combining two-view constraints for motion estimation. InProceedings of the 2001 IEEE computer society conference on computer vision and pattern recognition. CVPR 2001, volume 2, pages II–II. IEEE, 2001

    2001

  19. [19]

    Rotation averaging.International journal of computer vision, 103(3):267–305, 2013

    Richard Hartley, Jochen Trumpf, Yuchao Dai, and Hongdong Li. Rotation averaging.International journal of computer vision, 103(3):267–305, 2013

    2013

  20. [20]

    Detector- free structure from motion

    Xingyi He, Jiaming Sun, Yifan Wang, Sida Peng, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Detector- free structure from motion. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21594–21603, 2024

    2024

  21. [21]

    Deepmvs: Learning multi-view stereopsis

    Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view stereopsis. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018

    2018

  22. [22]

    Optimal transport aggregation for visual place recognition

    Sergio Izquierdo and Javier Civera. Optimal transport aggregation for visual place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2024

    2024

  23. [23]

    Mapanything: Universal feed-forward metric 3d reconstruction.arXiv preprint arXiv:2509.13414, 2025

    Nikhil Keetha, Norman Müller, Johannes Schönberger, Lorenzo Porzi, Yuchen Zhang, Tobias Fischer, Arno Knapitsch, Duncan Zauss, Ethan Weber, Nelson Antunes, et al. Mapanything: Universal feed-forward metric 3d reconstruction.arXiv preprint arXiv:2509.13414, 2025

    Pith/arXiv arXiv 2025

  24. [24]

    Tanks and temples: Benchmarking large-scale scene reconstruction.ACM Transactions on Graphics (ToG), 36(4):1–13, 2017

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

    2017

  25. [25]

    Rago: Recurrent graph optimizer for multiple rota- tion averaging

    Heng Li, Zhaopeng Cui, Shuaicheng Liu, and Ping Tan. Rago: Recurrent graph optimizer for multiple rota- tion averaging. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15787–15796, 2022

    2022

  26. [26]

    Matrixcity: A large-scale city dataset for city-scale neural rendering and beyond

    Yixuan Li, Lihan Jiang, Linning Xu, Yuanbo Xiangli, Zhenzhi Wang, Dahua Lin, and Bo Dai. Matrixcity: A large-scale city dataset for city-scale neural rendering and beyond. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 3205–3215, 2023

    2023

  27. [27]

    Megadepth: Learning single-view depth prediction from internet photos

    Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041–2050, 2018

    2041

  28. [28]

    Wint3r: Window-based streaming reconstruction with camera token pool

    Zizun Li, Jianjun Zhou, Yifan Wang, Haoyu Guo, Wenzheng Chang, Yang Zhou, Haoyi Zhu, Junyi Chen, Chunhua Shen, and Tong He. Wint3r: Window-based streaming reconstruction with camera token pool. arXiv preprint arXiv:2509.05296, 2025

    Pith/arXiv arXiv 2025

  29. [29]

    Chen, Zhenyu Li, Guang Shi, Jiashi Feng, and Bingyi Kang

    Haotong Lin, Sili Chen, Jun Hao Liew, Donny Y . Chen, Zhenyu Li, Guang Shi, Jiashi Feng, and Bingyi Kang. Depth anything 3: Recovering the visual space from any views.arXiv preprint arXiv:2511.10647, 2025. 11

    Pith/arXiv arXiv 2025

  30. [30]

    Lightglue: Local feature matching at light speed.arXiv preprint arXiv:2306.13643, 2023

    Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed.arXiv preprint arXiv:2306.13643, 2023

    Pith/arXiv arXiv 2023

  31. [31]

    Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision

    Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024

    2024

  32. [32]

    Vggt-slam 2.0: Real-time dense feed-forward scene reconstruction

    Dominic Maggio and Luca Carlone. Vggt-slam 2.0: Real-time dense feed-forward scene reconstruction. arXiv preprint arXiv:2601.19887, 2026

    arXiv 2026

  33. [33]

    Vggt-slam: Dense rgb slam optimized on the sl (4) manifold.arXiv preprint arXiv:2505.12549, 2025

    Dominic Maggio, Hyungtae Lim, and Luca Carlone. Vggt-slam: Dense rgb slam optimized on the sl (4) manifold.arXiv preprint arXiv:2505.12549, 2025

    Pith/arXiv arXiv 2025

  34. [34]

    Robust rotation and translation estimation in multiview reconstruction

    Daniel Martinec and Tomas Pajdla. Robust rotation and translation estimation in multiview reconstruction. In2007 IEEE conference on computer vision and pattern recognition, pages 1–8. IEEE, 2007

    2007

  35. [35]

    Nerf: Representing scenes as neural radiance fields for view synthesis.Communications of the ACM, 65 (1):99–106, 2021

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

    2021

  36. [36]

    Mast3r-slam: Real-time dense slam with 3d reconstruction priors

    Riku Murai, Eric Dexheimer, and Andrew J Davison. Mast3r-slam: Real-time dense slam with 3d reconstruction priors. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 16695–16705, 2025

    2025

  37. [37]

    Global Structure-from-Motion Revisited

    Linfei Pan, Daniel Barath, Marc Pollefeys, and Johannes Lutz Schönberger. Global Structure-from-Motion Revisited. InEuropean Conference on Computer Vision (ECCV), 2024

    2024

  38. [38]

    Aria digital twin: A new benchmark dataset for egocentric 3d machine perception

    Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Yuheng Carl Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 20133–20143, 2023

    2023

  39. [39]

    Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction

    Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. InInternational Conference on Computer Vision, 2021

    2021

  40. [40]

    Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding

    Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. InProceedings of the IEEE/CVF international conference on computer vision, pages 10912–10922, 2021

    2021

  41. [41]

    Superglue: Learning feature matching with graph neural networks

    Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020

    2020

  42. [42]

    Structure-from-motion revisited

    Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 4104–4113, 2016

    2016

  43. [43]

    Structure-from-motion revisited

    Johannes Lutz Schönberger and Jan-Michael Frahm. Structure-from-motion revisited. InConference on Computer Vision and Pattern Recognition (CVPR), 2016

    2016

  44. [44]

    Schönberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger

    Thomas Schöps, Johannes L. Schönberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi- camera videos. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2017

    2017

  45. [45]

    Fastvggt: Training-free acceleration of visual geometry transformer.arXiv preprint arXiv:2509.02560, 2025

    You Shen, Zhipeng Zhang, Yansong Qu, Xiawu Zheng, Jiayi Ji, Shengchuan Zhang, and Liujuan Cao. Fastvggt: Training-free acceleration of visual geometry transformer.arXiv preprint arXiv:2509.02560, 2025

    Pith/arXiv arXiv 2025

  46. [46]

    Flowmap: High-quality camera poses, intrinsics, and depth via gradient descent.arXiv preprint arXiv:2404.15259, 2024

    Cameron Smith, David Charatan, Ayush Tewari, and Vincent Sitzmann. Flowmap: High-quality camera poses, intrinsics, and depth via gradient descent.arXiv preprint arXiv:2404.15259, 2024

    Pith/arXiv arXiv 2024

  47. [47]

    Sturm, N

    J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers. A benchmark for the evaluation of rgb-d slam systems. InProc. of the International Conference on Intelligent Robot Systems (IROS), Oct. 2012

    2012

  48. [48]

    Loftr: Detector-free local feature matching with transformers

    Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922–8931, 2021. 12

    2021

  49. [49]

    Nerfstudio: A modular framework for neural radiance field development

    Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David McAllister, and Angjoo Kanazawa. Nerfstudio: A modular framework for neural radiance field development. InACM SIGGRAPH 2023 Conference Proceedings, SIGGRAPH ’23, 2023

    2023

  50. [50]

    Ba-net: Dense bundle adjustment network.arXiv preprint arXiv:1806.04807, 2018

    Chengzhou Tang and Ping Tan. Ba-net: Dense bundle adjustment network.arXiv preprint arXiv:1806.04807, 2018

    Pith/arXiv arXiv 2018

  51. [51]

    Deepv2d: Video to depth with differentiable structure from motion.arXiv preprint arXiv:1812.04605, 2018

    Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion.arXiv preprint arXiv:1812.04605, 2018

    Pith/arXiv arXiv 2018

  52. [52]

    DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras.Advances in neural information processing systems, 2021

    Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras.Advances in neural information processing systems, 2021

    2021

  53. [53]

    Smd-nets: Stereo mixture density networks

    Fabio Tosi, Yiyi Liao, Carolin Schmitt, and Andreas Geiger. Smd-nets: Stereo mixture density networks. InConference on Computer Vision and Pattern Recognition (CVPR), 2021

    2021

  54. [54]

    Bundle adjustment—a modern synthesis

    Bill Triggs, Philip F McLauchlan, Richard I Hartley, and Andrew W Fitzgibbon. Bundle adjustment—a modern synthesis. InInternational workshop on vision algorithms, pages 298–372. Springer, 1999

    1999

  55. [55]

    Amb3r: Accurate feed-forward metric-scale 3d reconstruction with backend.arXiv preprint arXiv:2511.20343, 2025

    Hengyi Wang and Lourdes Agapito. Amb3r: Accurate feed-forward metric-scale 3d reconstruction with backend.arXiv preprint arXiv:2511.20343, 2025

    arXiv 2025

  56. [56]

    Vggsfm: Visual geometry grounded deep structure from motion

    Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. Vggsfm: Visual geometry grounded deep structure from motion. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21686–21697, 2024

    2024

  57. [57]

    Vggt: Visual geometry grounded transformer

    Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 5294–5306, 2025

    2025

  58. [58]

    Continuous 3d perception model with persistent state

    Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 10510–10522, 2025

    2025

  59. [59]

    Dust3r: Geometric 3d vision made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697–20709, 2024

    2024

  60. [60]

    Tartanair: A dataset to push the limits of visual slam

    Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909–4916. IEEE, 2020

    2020

  61. [61]

    π3: Scalable permutation-equivariant visual geometry learning, 2025

    Yifan Wang, Jianjun Zhou, Haoyi Zhu, Wenzheng Chang, Yang Zhou, Zizun Li, Junyi Chen, Jiangmiao Pang, Chunhua Shen, and Tong He. π3: Scalable permutation-equivariant visual geometry learning, 2025. URLhttps://arxiv.org/abs/2507.13347

    Pith/arXiv arXiv 2025

  62. [62]

    Deepsfm: Structure from motion via deep bundle adjustment

    Xingkui Wei, Yinda Zhang, Zhuwen Li, Yanwei Fu, and Xiangyang Xue. Deepsfm: Structure from motion via deep bundle adjustment. InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I 16, pages 230–247. Springer, 2020

    2020

  63. [63]

    Robust global translations with 1dsfm

    Kyle Wilson and Noah Snavely. Robust global translations with 1dsfm. InEuropean conference on computer vision, pages 61–75. Springer, 2014

    2014

  64. [64]

    Point3r: Streaming 3d reconstruction with explicit spatial pointer memory.arXiv preprint arXiv:2507.02863, 2025

    Yuqi Wu, Wenzhao Zheng, Jie Zhou, and Jiwen Lu. Point3r: Streaming 3d reconstruction with explicit spatial pointer memory.arXiv preprint arXiv:2507.02863, 2025

    arXiv 2025

  65. [65]

    Rgbd objects in the wild: Scaling real-world 3d object learning from rgb-d videos

    Hongchi Xia, Yang Fu, Sifei Liu, and Xiaolong Wang. Rgbd objects in the wild: Scaling real-world 3d object learning from rgb-d videos. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22378–22389, 2024

    2024

  66. [66]

    Scal3r: Scalable test-time training for large-scale 3d reconstruction, 2026

    Tao Xie, Peishan Yang, Yudong Jin, Yingfeng Cai, Wei Yin, Weiqiang Ren, Qian Zhang, Wei Hua, Sida Peng, Xiaoyang Guo, and Xiaowei Zhou. Scal3r: Scalable test-time training for large-scale 3d reconstruction, 2026. URLhttps://arxiv.org/abs/2604.08542

    Pith/arXiv arXiv 2026

  67. [67]

    Vggt-motion: Motion-aware calibration-free monocular slam for long-range consistency.arXiv preprint arXiv:2602.05508, 2026

    Zhuang Xiong, Chen Zhang, Qingshan Xu, and Wenbing Tao. Vggt-motion: Motion-aware calibration-free monocular slam for long-range consistency.arXiv preprint arXiv:2602.05508, 2026. 13

    arXiv 2026

  68. [68]

    360recon: An accurate reconstruction method based on depth fusion from 360 images

    Zhongmiao Yan, Qi Wu, Songpengcheng Xia, Junyuan Deng, Xiang Mu, Renbiao Jin, Changchun Ye, and Ling Pei. 360recon: An accurate reconstruction method based on depth fusion from 360 images. In2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 17556–17563. IEEE, 2025

    2025

  69. [69]

    Blendedmvs: A large-scale dataset for generalized multi-view stereo networks.Computer Vision and Pattern Recognition (CVPR), 2020

    Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks.Computer Vision and Pattern Recognition (CVPR), 2020

    2020

  70. [70]

    Scannet++: A high-fidelity dataset of 3d indoor scenes

    Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 12–22, 2023

    2023

  71. [71]

    Bundle adjustment in the eager mode.IEEE Transactions on Robotics, 2026

    Zitong Zhan, Huan Xu, Zihang Fang, Xinpeng Wei, Yaoyu Hu, and Chen Wang. Bundle adjustment in the eager mode.IEEE Transactions on Robotics, 2026. URL https://arxiv.org/abs/2409.12190

    Pith/arXiv arXiv 2026

  72. [72]

    Loger: Long-context geometric reconstruction with hybrid memory.arXiv preprint arXiv:2603.03269, 2026

    Junyi Zhang, Charles Herrmann, Junhwa Hur, Chen Sun, Ming-Hsuan Yang, Forrester Cole, Trevor Darrell, and Deqing Sun. Loger: Long-context geometric reconstruction with hybrid memory.arXiv preprint arXiv:2603.03269, 2026

    Pith/arXiv arXiv 2026

  73. [73]

    Instantsfm: Towards gpu-native sfm for the deep learning era.arXiv preprint arXiv:2510.13310, 2025

    Jiankun Zhong, Zitong Zhan, Quankai Gao, Ziyu Chen, Haozhe Lou, Jiageng Mao, Ulrich Neumann, Chen Wang, and Yue Wang. Instantsfm: Towards gpu-native sfm for the deep learning era.arXiv preprint arXiv:2510.13310, 2025. URLhttps://arxiv.org/abs/2510.13310

    arXiv 2025

  74. [74]

    Omniworld: A multi-domain and multi-modal dataset for 4d world modeling.arXiv preprint arXiv:2509.12201, 2025

    Yang Zhou, Yifan Wang, Jianjun Zhou, Wenzheng Chang, Haoyu Guo, Zizun Li, Kaijing Ma, Xinyue Li, Yating Wang, Haoyi Zhu, et al. Omniworld: A multi-domain and multi-modal dataset for 4d world modeling.arXiv preprint arXiv:2509.12201, 2025. 14 In this appendix, we provide additional details and results to complement the main paper. Sec. A describes the netw...

    arXiv 2025

  75. [75]

    This adjustment is only applied to ensure stable inference on these challenging cases, and does not change the overall pipeline or optimization procedure. GPU-based optimization.The global positioning and bundle adjustment stages can become a computational bottleneck if implemented with conventional CPU-based solvers, especially for long sequences with ma...

This paper was first reviewed by grok-4.5 on July 13, 2026.