Pith. sign in

REVIEW 5 major objections 5 minor 50 references

By anchoring deformable Gaussian splatting to multi-view-consistent depth from a pretrained 3D transformer, Ground4D reconstructs dynamic 4D scenes from a single monocular video with the best reported geometry and rendering accuracy on the

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 09:44 UTC pith:442WMJB2

load-bearing objection A solid incremental pipeline for combining VGGT initialization with dynamic Gaussian refinement; the geometry claim is plausible but needs dynamic-region metrics to rule out self-anchoring. the 5 major comments →

arxiv 2606.28828 v2 pith:442WMJB2 submitted 2026-06-27 cs.CV

Ground4D: Consistency-Aware 4D Reconstruction from Monocular Video

Qing Zhao , Weijian Deng , Pengxu Wei , Liang Lin This is my paper
classification cs.CV
keywords 4D reconstructiondynamic Gaussian splattingmonocular videogeometry consistencyfoundation modelsnovel view synthesisdepth supervisioncontinuous-time dynamics
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 tries to establish that dynamic Gaussian splatting, which optimizes only photometric fidelity, drifts structurally, and that grounding it in geometry from a 3D foundation model—first as initialization, then as persistent depth supervision at observed and virtual viewpoints—fixes that drift. If true, a single monocular video yields both faithful 4D geometry and photorealistic novel views at arbitrary timestamps. The authors report a mean Chamfer distance of 0.226 and mPSNR of 19.43 dB with ground-truth poses and 19.07 dB without, outperforming prior dynamic Gaussian and foundation-model methods. Their ablations show the gains come from the geometry-consistency loss, especially on synthesized viewpoints, and from masking dynamic tokens in shallow attention layers during initialization.

Core claim

A monocular dynamic scene can be reconstructed in 4D by combining VGGT's multi-view-consistent point maps, produced training-free with dynamic tokens suppressed in shallow global attention, with a deformable Gaussian Splatting model refined under a geometry-consistency loss. Because VGGT predicts all point maps in one shared world frame, projecting the static plus per-frame dynamic points into any camera—including perturbed virtual cameras—yields depth targets that are consistent across viewpoints; aligning rendered depth to these targets anchors the Gaussians and preserves structure in regions not constrained by the image reconstruction loss. Continuous B-spline interpolation of the deforma

What carries the argument

The load-bearing object is VGGT's shared-world-frame point map: static points aggregated across frames plus per-frame dynamic points, projected to produce multi-view-consistent depth. The geometry-consistency loss compares rendered depth with this projected depth at both training views and perturbed virtual cameras, so all depth supervision derives from a single 3D structure. A training-free masking of dynamic tokens in the first five global attention layers keeps that structure clean, and B-spline node trajectories lift the discrete deformation field to continuous time.

Load-bearing premise

The per-frame dynamic point clouds predicted by VGGT—a static-scene model—are accurate enough to serve as geometric anchors; if they are wrong for moving objects, the geometry-consistency loss pulls the Gaussians toward incorrect structure.

What would settle it

Measure geometry error only inside dynamic regions, using ground-truth masks and a depth or multi-view ground truth, and compare Ground4D against a photometric-only dynamic Gaussian baseline. If dynamic-region Chamfer distance or accuracy does not improve—or worsens—the claimed geometry grounding is not actually helping where it matters.

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

If this is right

  • A single monocular video becomes enough to produce both accurate 4D geometry (mean Chamfer 0.226) and photorealistic novel views (mPSNR 19.43 dB with poses, 19.07 dB without) on DyCheck.
  • Because the depth anchors are projected from one shared 3D structure, the same geometry-consistency loss can be added to other dynamic Gaussian backbones; the ablation shows most of the gain comes from this loss, not from initialization alone.
  • Virtual-view supervision is cheap: even one perturbed camera per training view improves both rendering and geometry, and the result is insensitive to perturbation size, suggesting easy transfer to other settings.
  • Continuous B-spline node trajectories let the model render and reconstruct at arbitrary timestamps, including extrapolated past and future times, which discrete-frame methods cannot do.
  • Estimated camera trajectories also improve (ATE 0.010 on DyCheck), so the pipeline can operate without ground-truth poses, which matters for real-world deployment.

Where Pith is reading between the lines

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

  • The paper's final paragraph acknowledges computational cost for long sequences and dependence on SAM masks; it does not acknowledge a deeper risk—VGGT is a static-scene model, and its per-frame dynamic points are not temporally consistent—and the ablations do not report geometry accuracy restricted to dynamic regions. The dynamic-part geometry claim is therefore untested.
  • If the VGGT-derived geometry is systematically biased, the method inherits that bias; a stress test on sequences with fast motion, occlusions, or reflective surfaces would reveal where photometric-only optimization might actually be safer.
  • A natural extension is to replace external masks with motion cues already inside the transformer, which the authors list as future work; since their mask-source ablation shows little sensitivity, such intrinsic masks could remove the SAM dependency without hurting accuracy.
  • The consistent-depth-supervision recipe transfers beyond this paper: any module that outputs a shared-world-frame point cloud could initialize and anchor a Gaussian or radiance field, potentially making this a general pattern for dynamic reconstruction.

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

5 major / 5 minor

Summary. The paper proposes Ground4D, a two-stage method for dynamic 4D reconstruction from monocular video. The first stage uses VGGT, a pretrained 3D foundation model, in a training-free manner: dynamic tokens are masked in early global attention layers to obtain camera poses, a global static point cloud, and per-frame dynamic point clouds. The second stage initializes a deformable 3D Gaussian Splatting model from these point clouds and optimizes it with a photometric loss, a geometry-consistency loss that aligns rendered depth with VGGT-projected depth at both training and synthetic virtual viewpoints, and motion regularization losses. Experiments on DyCheck report state-of-the-art geometry (Chamfer distance 0.226) and novel-view synthesis (mPSNR 19.43 with GT poses, 19.07 without GT poses), with ablations on masking layers, refinement, mask source, and virtual-view sampling. The paper also claims continuous 4D dynamics via B-spline interpolation of deformation nodes, shown qualitatively on DAVIS.

Significance. If the results hold, Ground4D is a useful integration of geometry priors from a foundation model into dynamic Gaussian optimization. The paper is clearly written and the component-wise ablations, especially the training-free dynamic token masking and the virtual-view geometry-consistency loss, are well designed and support the individual claims. The experiments are internally consistent and the external ground-truth benchmark mitigates the most extreme form of circularity. However, the geometric supervision is derived from the same VGGT model used for initialization, and the paper does not evaluate dynamic-region geometry separately, which is the key weakness. The continuous-time rendering claim is also only qualitative. These issues need to be addressed before the central claims can be fully accepted.

major comments (5)
  1. [§III-C, Eq. (2)] Equation (2) defines the geometry-consistency loss as a per-pixel L1 distance, but the text states that the depth supervision is applied in a 'scale-invariant manner.' The exact normalization/alignment is never specified. Since rendered depth and VGGT-projected depth generally live in different units, the actual loss is ambiguous and not reproducible. Please provide the precise formulation, including how the scale is estimated (per-frame median? similarity transform?) and how pixels with missing projected depth are masked.
  2. [§III-B/III-C, Tables I and VI] The load-bearing risk is that the per-frame dynamic point clouds P_dyn_t are obtained from a static-scene model (VGGT) by masking dynamic tokens, and these same points are used both to initialize dynamic Gaussians and to supervise the geometry-consistency loss. If P_dyn_t is biased — e.g., wrong depth or fragmented dynamic regions — the refinement anchors Gaussians to that bias. The current experiments do not measure dynamic-region geometry: Table VI only reports aggregate mean Accuracy/Completeness, and Table I reports aggregate Chamfer distances, both dominated by static background. The manuscript's own limitation statement (§V) acknowledges reliance on the foundation model, but this is not tested. Please report dynamic-only Accuracy/Completeness/Chamfer (e.g., using DyCheck ground-truth dynamic masks/trajectories) and, if feasible, independently evaluate VGGT's dynamic point clouds ag
  3. [§III-C and Fig. 5] The abstract and Section III claim that Ground4D 'inherently models continuous 4D dynamics' and supports rendering at arbitrary timestamps. However, the only evidence is qualitative visualization in Fig. 5. No quantitative metrics are reported for interpolated or extrapolated timestamps, no comparison is made to other continuous-time baselines, and temporal smoothness is not measured. Since this is stated as a capability of the method, please add a quantitative experiment on DyCheck (e.g., held-out timestamp interpolation with mPSNR/mSSIM and Chamfer) or explicitly soften the claim to a qualitative demonstration.
  4. [§IV-B, Table II] Table II mixes comparison protocols. The caption states that Shape-of-Motion uses a 5-scene subset at 1x resolution, while all other methods are evaluated on 7 scenes at 2x resolution. This makes the reported SOTA comparisons (e.g., 19.43 mPSNR vs 17.32 for Shape-of-Motion) not directly interpretable. Please either report all methods on the same scenes and resolution or separate the three protocols into distinct tables with clear labeling.
  5. [§IV-A/IV-C, Table VII] Hyperparameters (L_s, lambda_s, lambda_gc, lambda_r, K, perturbation range) are selected on the DyCheck benchmark itself, and no repeated-seed statistics are provided. Given that Table VII shows only small differences across virtual-view configurations, the claims would be materially strengthened by reporting mean±std over several optimization seeds and by stating how hyperparameters were chosen (e.g., a separate validation split). This is a standard reproducibility requirement for optimization-based scene reconstruction.
minor comments (5)
  1. [§III-B, Eq. (1)] The notation 't^{(l)}_t(k)' is confusing: the subscript t is reused for both the token index and the frame index. Please use distinct symbols for tokens and frames.
  2. [§III-A] The paper states that per-frame dynamic masks can be obtained via SAM but does not specify how SAM is applied to video (prompting strategy, instance tracking, or per-frame segmentation). Please provide details.
  3. [§IV-A] The geometry evaluation protocol is described only as following prior work [9]. Please specify how ground-truth point clouds are sampled, whether a similarity transform is used to align predictions, and how median/mean metrics are computed.
  4. [§V] The 'Potential limitations' paragraph acknowledges computational cost and reliance on SAM, but no quantitative analysis of runtime or failure cases is given. Including such analysis would improve the paper.
  5. [General] No code or model release is mentioned. For reproducibility of an optimization-based method with many hyperparameters, code release is strongly recommended.

Circularity Check

0 steps flagged

No significant circularity: reported results are evaluated against external DyCheck ground truth, and the VGGT-derived depth supervision is a regularizer rather than the evaluation target.

full rationale

Ground4D's derivation chain is not circular. Stage 1 (Sec. III-B) obtains VGGT point clouds and camera poses; Stage 2 (Sec. III-C) initializes Gaussians from those points and adds L_gc (Eq. 2), which aligns rendered depth to depth projected from the same VGGT structure. This shared source is a robustness limitation: if VGGT's per-frame dynamic point clouds are biased, L_gc can anchor Gaussians to that bias, and the paper does not report per-region dynamic geometry accuracy. But this is a correctness/evidence gap, not a circularity. The headline geometry metrics are Accuracy/Completeness/Chamfer against external DyCheck ground truth, not against VGGT depth, and the photometric loss L_rgb plus ARAP/velocity/acceleration regularizers can move Gaussians away from the initialization. The paper does not define its output as the VGGT input; it uses VGGT as a prior and evaluates against independent ground truth. No self-citation is load-bearing: VGGT [10], SAM [40], and DyCheck [45] are external works, and the paper's ablations (Tables IV–VII) test the added components empirically. The skeptical concern about unmeasured dynamic-region accuracy should be recorded as a validation limitation, not as circular reasoning.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claim rests on the accuracy and consistency of VGGT geometry, the correctness of SAM masks, and the validity of the deformation-graph model adopted from MoSca. The hand-set loss weights and ablation-selected hyperparameters (L_s, K, perturbation ranges) are tuned on the evaluation benchmark, which increases the risk of overfitting the reported metrics. No fundamentally new physical or model entities are introduced.

free parameters (6)
  • lambda_s = 0.5
    Weight for the virtual-view depth consistency loss in Eq. 2; hand-set, not from theory.
  • lambda_gc = 0.1
    Weight for the geometry-consistency loss in Eq. 3; hand-set.
  • lambda_r = 0.01
    Weight for ARAP/velocity/acceleration losses in Eq. 3; hand-set.
  • L_s = 5
    Number of shallow global attention layers with dynamic-token key/value masking; chosen by ablation on DyCheck (Table IV).
  • K = 1
    Number of virtual cameras sampled per training view; chosen by ablation on DyCheck (Table VII).
  • perturbation_range = rot 5 deg, trans 0.1
    Rotation/translation perturbation magnitudes for virtual cameras; chosen by ablation on DyCheck (Table VII).
axioms (5)
  • domain assumption VGGT predicts point maps for all frames in a shared world coordinate system (first camera frame), so projecting any camera viewpoint yields multi-view-consistent depth.
    Invoked in Sec III-B ('A key property...') and used to define \hat D_t and \hat D_{t,k}; if this property degrades under token masking, the consistency supervision is invalid.
  • domain assumption SAM-provided dynamic masks correctly identify all moving regions; missed dynamic pixels are treated as static and corrupt geometry.
    Sec III-A says masks are obtained via SAM; the dynamic token masking relies on mask accuracy, and the paper's own limitation section acknowledges this dependency.
  • ad hoc to paper Masking dynamic tokens' key/value in the first L_s=5 global attention layers suppresses motion interference without a harmful distribution shift.
    Empirical heuristic validated only by ablation on DyCheck (Table IV); it is not derived from theory.
  • domain assumption VGGT-estimated camera poses are accurate enough that projecting points into virtual cameras gives valid geometric supervision.
    Virtual views are synthesized by perturbing estimated cameras (Sec III-C); pose errors propagate directly into the depth targets.
  • domain assumption MoSca's deformation-graph representation with SE(3) node trajectories can express the scene motion, including the ARAP/velocity/acceleration priors.
    Sec III-C adopts the deformation field from [5] without re-derivation; the regularization losses assume local rigidity and smoothness that may not hold for all dynamic scenes.

pith-pipeline@v1.3.0-alltime-deepseek · 15230 in / 13231 out tokens · 118204 ms · 2026-08-02T09:44:00.111751+00:00 · methodology

0 comments
read the original abstract

Learning a 4D scene representation from a single monocular video that supports dynamic novel-view synthesis while maintaining faithful geometry over time remains challenging. Dynamic Gaussian Splatting achieves strong rendering performance through photometric optimization, yet does not explicitly enforce multi-view geometric consistency. In contrast, 3D foundation models recover coherent scene geometry and camera motion, but their point-based outputs are not designed for photorealistic rendering. We propose Ground4D, a geometry-grounded framework built on two stages. First, we perform geometry initialization via 3D foundation models, leveraging VGGT in a training-free manner to reconstruct multi-view-consistent 3D geometry and camera poses from monocular video. The recovered geometry provides a structured and reliable initialization for dynamic Gaussian representations. Second, we conduct geometry-consistency-aware refinement via dynamic Gaussian Splatting, optimizing the representation through differentiable rendering while maintaining multi-view geometric consistency across both observed and synthesized viewpoints. Furthermore, Ground4D inherently models the continuous 4D dynamics of the scene, naturally supporting rendering at arbitrary timestamps. By integrating foundation-level geometric priors into dynamic Gaussian optimization, Ground4D achieves stronger reconstruction fidelity and rendering performance, underscoring the role of geometry-grounded constraints in robust 4D scene modeling.

Figures

Figures reproduced from arXiv: 2606.28828 by Liang Lin, Pengxu Wei, Qing Zhao, Weijian Deng.

Figure 1
Figure 1. Figure 1: Ground4D enables consistency-aware 4D geometry reconstruction and high-quality novel-view synthesis from a single monocular video. While direct outputs from 3d foundation models like VGGT [10] provide strong spatial priors, they lack the fidelity for dynamic rendering. To bridge this gap, we first perform a training-free 4D geometry initialization to yield a geometrically grounded starting representation. … view at source ↗
Figure 2
Figure 2. Figure 2: Ground4D Overview. Given a monocular video and corresponding dynamic masks extracted via SAM [40], our method proceeds in two stages. (1) Geometry Initialization via 3D Foundation Models: We leverage VGGT in a training-free manner to recover multi-view-consistent 3D geometry and camera parameters by suppressing dynamic tokens within the global attention layers. The recovered 3D points initialize dynamic Ga… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative Results of the Reconstructed 4D Geometry. Our Ground4D achieves geometry consistent reconstruction of both static scenes and moving objects. TABLE III QUANTITATIVE COMPARISONS OF CAMERA POSE ESTIMATION ON THE DYCHECK [45] AND TUM-DYNAMICS [48] DATASET. DyCheck [45] TUM-dynamics [48] Method ATE ↓ RTE ↓ RRE ↓ ATE ↓ RTE ↓ RRE ↓ DUSt3R [7] 0.035 0.030 2.323 0.100 0.087 2.692 CUT3R [30] 0.029 0.020 … view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative results of novel view synthesis. Ground4D effectively eliminates the blurring and artifacts, yielding sharper and more photorealistic novel views. Continuous 4D Dynamics Continuous 4D Dynamics Past: T = -0.1 Between: T = 0.1 Between: T = 0.1 Between: T = 0.2 Future: T = 1.1 Between: T = 0.3 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of continuous-time novel-view synthesis. The temporal range of each input video is normalized to [0, 1]. Ground4D queries the learned continuous 4D dynamics at arbitrary timestamps to render dynamic scenes from novel viewpoints. The results show smooth foreground motion and coherent scene geometry, demonstrating that Ground4D lifts discrete frame-wise 4D reconstruction into continuous-time dy… view at source ↗

Review history (2 revisions) →

discussion (0)

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

Reference graph

Works this paper leans on

50 extracted references · 2 linked inside Pith

  1. [1]

    4d gaussian splatting for real-time dynamic scene rendering,

    G. Wu, T. Yi, J. Fang, L. Xie, X. Zhang, W. Wei, W. Liu, Q. Tian, and X. Wang, “4d gaussian splatting for real-time dynamic scene rendering,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 20 310–20 320

    2024

  2. [2]

    Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction,

    Z. Yang, X. Gao, W. Zhou, S. Jiao, Y . Zhang, and X. Jin, “Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 20 331–20 341

    2024

  3. [3]

    Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis,

    J. Luiten, G. Kopanas, B. Leibe, and D. Ramanan, “Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis,” in2024 International Conference on 3D Vision (3DV). IEEE, 2024, pp. 800– 809

    2024

  4. [4]

    Dynamic gaussian marbles for novel view synthesis of casual monocular videos,

    C. Stearns, A. Harley, M. Uy, F. Dubost, F. Tombari, G. Wetzstein, and L. Guibas, “Dynamic gaussian marbles for novel view synthesis of casual monocular videos,” inSIGGRAPH Asia 2024 Conference Papers, 2024, pp. 1–11

    2024

  5. [5]

    Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds,

    J. Lei, Y . Weng, A. W. Harley, L. Guibas, and K. Daniilidis, “Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds,” inProceedings of the Computer Vision and Pattern Recognition Confer- ence, 2025, pp. 6165–6177

    2025

  6. [6]

    Shape of motion: 4d reconstruction from a single video,

    Q. Wang, V . Ye, H. Gao, W. Zeng, J. Austin, Z. Li, and A. Kanazawa, “Shape of motion: 4d reconstruction from a single video,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 9660–9672. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 10

    2025

  7. [7]

    Dust3r: Geometric 3d vision made easy,

    S. Wang, V . Leroy, Y . Cabon, B. Chidlovskii, and J. Revaud, “Dust3r: Geometric 3d vision made easy,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 20 697–20 709

    2024

  8. [8]

    Monst3r: A simple approach for estimating geometry in the presence of motion,

    J. Zhang, C. Herrmann, J. Hur, V . Jampani, T. Darrell, F. Cole, D. Sun, and M.-H. Yang, “Monst3r: A simple approach for estimating geometry in the presence of motion,”International Conference on Learning Representations, 2025

    2025

  9. [9]

    Easi3r: Estimating disentangled motion from dust3r without training,

    X. Chen, Y . Chen, Y . Xiu, A. Geiger, and A. Chen, “Easi3r: Estimating disentangled motion from dust3r without training,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 9158–9168

    2025

  10. [10]

    Vggt: Visual geometry grounded transformer,

    J. Wang, M. Chen, N. Karaev, A. Vedaldi, C. Rupprecht, and D. Novotny, “Vggt: Visual geometry grounded transformer,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 5294– 5306

    2025

  11. [11]

    Page-4d: Disentangled pose and geometry estimation for 4d perception,

    K. Zhou, Y . Wang, G. Chen, G. Beaudouin, F. Zhan, P. P. Liang, and M. Wang, “Page-4d: Disentangled pose and geometry estimation for 4d perception,” inInternational Conference on Learning Representations, 2026

    2026

  12. [12]

    Vggt4d: Mining motion cues in visual geometry transformers for 4d scene reconstruction,

    Y . Hu, C. Cheng, S. Yu, X. Guo, and H. Wang, “Vggt4d: Mining motion cues in visual geometry transformers for 4d scene reconstruction,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 414–424

    2026

  13. [13]

    Building rome in a day,

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

    2011

  14. [14]

    Self-calibration and metric reconstruction inspite of varying and unknown intrinsic camera parame- ters,

    M. Pollefeys, R. Koch, and L. V . Gool, “Self-calibration and metric reconstruction inspite of varying and unknown intrinsic camera parame- ters,”International journal of computer vision, vol. 32, no. 1, pp. 7–25, 1999

    1999

  15. [15]

    Visual modeling with a hand-held camera,

    M. Pollefeys, L. Van Gool, M. Vergauwen, F. Verbiest, K. Cornelis, J. Tops, and R. Koch, “Visual modeling with a hand-held camera,” International Journal of Computer Vision, vol. 59, no. 3, pp. 207–232, 2004

    2004

  16. [16]

    Structure-from-motion revisited,

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

    2016

  17. [17]

    Photo tourism: exploring photo collections in 3d,

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

    2006

  18. [18]

    Modeling the world from internet photo collections,

    ——, “Modeling the world from internet photo collections,”Interna- tional journal of computer vision, vol. 80, no. 2, pp. 189–210, 2008

    2008

  19. [19]

    Bundle adjustment in the large,

    S. Agarwal, N. Snavely, S. M. Seitz, and R. Szeliski, “Bundle adjustment in the large,” inEuropean conference on computer vision. Springer, 2010, pp. 29–42

    2010

  20. [20]

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

    E. Brachmann, J. Wynn, S. Chen, T. Cavallari, A. Monszpart, D. Tur- mukhambetov, and V . A. Prisacariu, “Scene coordinate reconstruction: Posing of image collections via incremental learning of a relocalizer,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 421– 440

    2024

  21. [21]

    Ba-net: Dense bundle adjustment network,

    C. Tang and P. Tan, “Ba-net: Dense bundle adjustment network,” International Conference on Learning Representations, 2019

    2019

  22. [22]

    Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras,

    Z. Teed and J. Deng, “Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras,”Advances in neural information processing systems, vol. 34, pp. 16 558–16 569, 2021

    2021

  23. [23]

    Vggsfm: Visual geometry grounded deep structure from motion,

    J. Wang, N. Karaev, C. Rupprecht, and D. Novotny, “Vggsfm: Visual geometry grounded deep structure from motion,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 21 686–21 697

    2024

  24. [24]

    Grounding image matching in 3d with mast3r,

    V . Leroy, Y . Cabon, and J. Revaud, “Grounding image matching in 3d with mast3r,” inEuropean conference on computer vision. Springer, 2024, pp. 71–91

    2024

  25. [25]

    Fast3r: Towards 3d reconstruction of 1000+ images in one forward pass,

    J. Yang, A. Sax, K. J. Liang, M. Henaff, H. Tang, A. Cao, J. Chai, F. Meier, and M. Feiszli, “Fast3r: Towards 3d reconstruction of 1000+ images in one forward pass,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 21 924–21 935

    2025

  26. [26]

    Streaming 4d visual geometry transformer,

    D. Zhuo, W. Zheng, J. Guo, Y . Wu, J. Zhou, and J. Lu, “Streaming 4d visual geometry transformer,”International Conference on Learning Representations, 2026

    2026

  27. [27]

    Infinitevggt: Visual geometry grounded transformer for endless streams,

    S. Yuan, Y . Yang, X. Yang, X. Zhang, Z. Zhao, L. Zhang, and Z. Zhang, “Infinitevggt: Visual geometry grounded transformer for endless streams,”arXiv preprint arXiv:2601.02281, 2026

    arXiv 2026

  28. [28]

    π 3: Permutation-equivariant visual geometry learning,

    Y . Wang, J. Zhou, H. Zhu, W. Chang, Y . Zhou, Z. Li, J. Chen, J. Pang, C. Shen, and T. He, “π 3: Permutation-equivariant visual geometry learning,”International Conference on Learning Representations, 2026

    2026

  29. [29]

    Fastvggt: Training-free acceleration of visual geometry transformer,

    Y . Shen, Z. Zhang, Y . Qu, X. Zheng, J. Ji, S. Zhang, and L. Cao, “Fastvggt: Training-free acceleration of visual geometry transformer,” International Conference on Learning Representations, 2026

    2026

  30. [30]

    Continuous 3d perception model with persistent state,

    Q. Wang, Y . Zhang, A. Holynski, A. A. Efros, and A. Kanazawa, “Continuous 3d perception model with persistent state,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 10 510–10 522

    2025

  31. [31]

    Das3r: Dynamics- aware gaussian splatting for static scene reconstruction,

    K. Xu, T. H. E. Tse, J. Peng, and A. Yao, “Das3r: Dynamics- aware gaussian splatting for static scene reconstruction,”arXiv preprint arXiv:2412.19584, 2024

    Pith/arXiv arXiv 2024

  32. [32]

    Vision transformers for dense prediction,

    R. Ranftl, A. Bochkovskiy, and V . Koltun, “Vision transformers for dense prediction,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 12 179–12 188

    2021

  33. [33]

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

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,”Communications of the ACM, vol. 65, no. 1, pp. 99–106, 2021

    2021

  34. [34]

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

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

    2021

  35. [35]

    Robust dynamic radiance fields,

    Y .-L. Liu, C. Gao, A. Meuleman, H.-Y . Tseng, A. Saraf, C. Kim, Y .-Y . Chuang, J. Kopf, and J.-B. Huang, “Robust dynamic radiance fields,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 13–23

    2023

  36. [36]

    Cbarf: cascaded bundle-adjusting neural radiance fields from imperfect camera poses,

    H. Fu, X. Yu, L. Li, and L. Zhang, “Cbarf: cascaded bundle-adjusting neural radiance fields from imperfect camera poses,”IEEE Transactions on Multimedia, vol. 26, pp. 9304–9315, 2024

    2024

  37. [37]

    4dgstream: Variable bitrate dynamic gaussian splatting streaming,

    Z. Liang, D. Zhang, L. Shen, M. Zhang, J. Zhang, B. Ju, M. Dasari, F. Wang, and J. Liu, “4dgstream: Variable bitrate dynamic gaussian splatting streaming,”IEEE Transactions on Multimedia, 2026

    2026

  38. [38]

    3d gaussian splatting for real-time radiance field rendering

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, G. Drettakiset al., “3d gaussian splatting for real-time radiance field rendering.”ACM Trans. Graph., vol. 42, no. 4, pp. 139–1, 2023

    2023

  39. [39]

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

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

    2024

  40. [40]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Loet al., “Segment anything,” inProceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 4015–4026

    2023

  41. [41]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Noubyet al., “Dinov2: Learning robust visual features without supervision,”Transactions on Machine Learning Research, 2024

    2024

  42. [42]

    Skinning with dual quaternions,

    L. Kavan, S. Collins, J. ˇZ´ara, and C. O’Sullivan, “Skinning with dual quaternions,” inProceedings of the 2007 symposium on Interactive 3D graphics and games, 2007, pp. 39–46

    2007

  43. [43]

    As-rigid-as-possible surface modeling,

    O. Sorkine, M. Alexaet al., “As-rigid-as-possible surface modeling,” in Symposium on Geometry processing, vol. 4, 2007, pp. 109–116

    2007

  44. [44]

    Dynamicfusion: Recon- struction and tracking of non-rigid scenes in real-time,

    R. A. Newcombe, D. Fox, and S. M. Seitz, “Dynamicfusion: Recon- struction and tracking of non-rigid scenes in real-time,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2015, pp. 343–352

    2015

  45. [45]

    Monocular dynamic view synthesis: A reality check,

    H. Gao, R. Li, S. Tulsiani, B. Russell, and A. Kanazawa, “Monocular dynamic view synthesis: A reality check,”Advances in Neural Informa- tion Processing Systems, vol. 35, pp. 33 768–33 780, 2022

    2022

  46. [46]

    A benchmark dataset and evaluation methodology for video object segmentation,

    F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung, “A benchmark dataset and evaluation methodology for video object segmentation,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 724–732

    2016

  47. [47]

    The 2017 davis challenge on video object segmen- tation,

    J. Pont-Tuset, F. Perazzi, S. Caelles, P. Arbel ´aez, A. Sorkine-Hornung, and L. Van Gool, “The 2017 davis challenge on video object segmen- tation,”arXiv preprint arXiv:1704.00675, 2017

    Pith/arXiv arXiv 2017

  48. [48]

    A benchmark for the evaluation of rgb-d slam systems,

    J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers, “A benchmark for the evaluation of rgb-d slam systems,” in2012 IEEE/RSJ international conference on intelligent robots and systems. IEEE, 2012, pp. 573–580

    2012

  49. [49]

    Unsupervised collaborative learning of keyframe detection and visual odometry towards monocular deep slam,

    L. Sheng, D. Xu, W. Ouyang, and X. Wang, “Unsupervised collaborative learning of keyframe detection and visual odometry towards monocular deep slam,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 4302–4311

    2019

  50. [50]

    Ttt3r: 3d recon- struction as test-time training,

    X. Chen, Y . Chen, Y . Xiu, A. Geiger, and A. Chen, “Ttt3r: 3d recon- struction as test-time training,”International Conference on Learning Representations, 2026

    2026