pith. sign in

arxiv: 2607.01885 · v1 · pith:PEZG2IZVnew · submitted 2026-07-02 · 💻 cs.CV

Diversity-aware View Partitioning for Scalable VGGT

Jinsoo Park , Donggyu Choi , Ahyun Seo , Minsu cho , Jeany Son This is my paper

Pith reviewed 2026-07-03 16:09 UTC · model grok-4.3

classification 💻 cs.CV
keywords view partitioninggeometry transformermulti-view reconstructionattention efficiencycamera pose estimation3D reconstructiongraph partitioningVGGT
0
0 comments X

The pith

Partitioning input views by visual dissimilarity and spatial dispersion lets VGGT scale to large collections with better accuracy and lower memory cost.

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

The paper establishes that VGGT and similar geometry transformers lose performance when given many redundant views because identical tokens dilute the attention signals required for accurate geometry. It therefore splits the full set into balanced chunks that each maximize viewpoint variety, using combinatorial graph partitioning on pairwise visual dissimilarity together with an approximation of spatial spread. The spread estimate comes from soft pose propagation seeded on a few frames and driven only by visual similarity, avoiding any full pose solve or model retraining. A sympathetic reader would care because the approach turns the quadratic attention bottleneck into a practical method for real-world multi-view datasets while preserving or improving results on pose, depth, and reconstruction.

Core claim

The central claim is that constructing diversity-aware balanced chunks through combinatorial graph partitioning over visual dissimilarity and approximated spatial dispersion allows the transformer to focus attention on geometrically informative views, reduces redundant interactions, and thereby improves camera pose estimation, multi-view depth prediction, and 3D reconstruction while lowering memory usage and inference latency, all in a training-free plug-and-play manner.

What carries the argument

Combinatorial graph partitioning over visual dissimilarity and spatial dispersion, with spatial dispersion obtained via soft pose propagation from a small set of seed frames using visual similarity.

If this is right

  • Camera pose estimation accuracy rises because attention operates on geometrically distinct rather than redundant views.
  • Multi-view depth prediction improves as geometric signals are no longer diluted by near-duplicate tokens.
  • 3D reconstruction quality increases while peak memory and inference latency decrease.
  • The same partitioning works as a drop-in addition to existing VGGT variants without retraining.
  • Collections of hundreds of views become computationally feasible without the quality drop previously observed from redundancy.

Where Pith is reading between the lines

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

  • The same partitioning idea could be tested on other global-attention multi-view architectures to check whether the benefit is specific to VGGT or general.
  • If the soft propagation proves reliable, it suggests that explicit pose estimation can sometimes be skipped for the narrower task of view selection.
  • Iteratively refining the seed set during capture might further tighten the dispersion estimate in large or complex scenes.
  • The method implies a possible online strategy: reject incoming frames whose visual similarity to existing chunks is too high.

Load-bearing premise

The soft pose propagation from a small set of seed frames based on visual similarity produces a sufficiently accurate approximation of spatial dispersion to guide effective partitioning.

What would settle it

Apply the partitioning on a dataset containing repetitive textures where the visual-similarity propagation visibly mismatches true camera positions, then compare reconstruction metrics against both random partitioning and the full unpartitioned baseline.

Figures

Figures reproduced from arXiv: 2607.01885 by Ahyun Seo, Donggyu Choi, Jeany Son, Jinsoo Park, Minsu Cho.

Figure 1
Figure 1. Figure 1: Effectiveness of diversity-aware view partitioning, (a) Our method reconstructs more geometrically consistent structures while preserving finer scene details compared to existing efficient VGGT variants, such as LiteVGGT [32]. (b) Unlike other meth￾ods [32, 40], whose performance tends to degrade as the number of input frames in￾creases, our simple diversity-aware view partitioning mitigates this effect an… view at source ↗
Figure 2
Figure 2. Figure 2: Impact of frame count on VGGT performance. Red boxes indicate high-error regions where VGGT struggles with many frames, while green boxes indicate low-error regions. Frames sampled from each region show that failures occur in visually redundant views, whereas successful regions exhibit greater viewpoint diversity. 3 Motivation In practical multi-view capture scenarios, frames are often densely sampled alon… view at source ↗
Figure 3
Figure 3. Figure 3: Attention entropy analysis. To better examine this behavior, we measure frame-level attention entropy un￾der different frame counts and sampling strategies. From the same sequences, we construct subsets of 300 and 500 frames, each sampled either redundantly (visually similar frames) or sparsely (diverse view￾points). As shown in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the proposed framework. Input frames are first partitioned using visual dissimilarity. A selected reference chunk is processed by VGGT to obtain pseudo￾spatial cues, which are used to estimate pseudo-poses for the remaining frames. The frames are then reorganized into balanced chunks that maximize both visual and spatial diversity, and each chunk is processed independently by VGGT. These findin… view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of Pseudo￾pose on NRGBD [1]. Pseudo-Pose Estimation [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visual–spatial correlation on 7Scenes [31]. Soft pose propagation (Sec. 4.3) is de￾signed as a coarse spatial cue for parti￾tioning rather than a metric pose esti￾mator. To assess the quality of this cue directly, we measure the median rela￾tive camera-center direction error for lo￾cal frame pairs on NRGBD [1], which is only 4.21◦ , confirming that the prop￾agated poses capture useful local spatial trends.… view at source ↗
Figure 7
Figure 7. Figure 7: 3D Point Cloud Reconstruction on NRGBD [1] and ScanNet-50 [7] [PITH_FULL_IMAGE:figures/full_fig_p029_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Results of 3D Point Cloud Reconstruction on NRGBD [1] [PITH_FULL_IMAGE:figures/full_fig_p030_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Results of 3D Point Cloud Reconstruction on ScanNet-50 [7] [PITH_FULL_IMAGE:figures/full_fig_p031_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Results of 3D Trajectories on ScanNet-50 [7] [PITH_FULL_IMAGE:figures/full_fig_p032_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Normal Consistency Comparison on 7Scenes [31] [PITH_FULL_IMAGE:figures/full_fig_p034_11.png] view at source ↗
read the original abstract

Geometry transformers such as VGGT achieve strong performance by jointly reasoning over multiple views with global attention. However, scaling them to large view collections remains challenging due to the quadratic cost of attention. Moreover, our empirical analysis reveals that the reconstruction quality in VGGT is sensitive to the distribution of viewpoints. Simply increasing the number of views without sufficient viewpoint diversity can even degrade performance, as redundant views introduce highly similar tokens that dilute informative geometric signals in the attention mechanism. Motivated by this observation, we propose a training-free and plug-and-play VGGT inference framework that organizes views into diversity-aware balanced chunks. The chunks are constructed through combinatorial graph partitioning over visual dissimilarity and spatial dispersion. This view organization allows the transformer to focus attention on geometrically informative views while reducing redundant attention interactions. To estimate spatial dispersion without full pose estimation, we approximate spatial relationships via a soft pose propagation strategy based on visual similarity from a small set of seed frames. Extensive experiments demonstrate improved performance in camera pose estimation, multi-view depth prediction, and 3D reconstruction while reducing memory usage and inference latency. Our framework also complements existing VGGT variants, enabling scalable multi-view reconstruction without sacrificing geometric fidelity.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes a training-free, plug-and-play inference framework for VGGT that partitions input views into diversity-aware balanced chunks via combinatorial graph partitioning over visual dissimilarity plus an approximated spatial dispersion term. The spatial dispersion is obtained by soft pose propagation from a small set of seed frames using only visual similarity. The claimed benefit is that the resulting chunks let the transformer focus attention on geometrically informative views, yielding gains in camera pose estimation, multi-view depth prediction, and 3D reconstruction together with lower memory and latency.

Significance. If the empirical results hold and the partitioning is shown to be robust, the work would provide a practical way to scale global-attention geometry transformers beyond small view counts without retraining, addressing both quadratic cost and the observed degradation from redundant viewpoints.

major comments (2)
  1. [Abstract / method description of soft pose propagation] The central claim that the diversity-aware chunks produce the reported gains rests on the soft pose propagation strategy yielding a sufficiently accurate proxy for spatial dispersion. The manuscript must supply ablations that compare the full method against (i) visual-similarity-only partitioning and (ii) random balanced chunks on the same backbones and datasets; without these controls it is impossible to attribute improvements to the spatial-dispersion term rather than to chunking in general.
  2. [Abstract / experimental section] No quantitative results, error bars, dataset statistics, or ablation tables appear in the abstract, and the provided text supplies none of the concrete numbers needed to evaluate the claimed improvements in pose estimation, depth, or reconstruction. The load-bearing experimental evidence must be presented with direct comparisons to the unpartitioned VGGT baseline.
minor comments (1)
  1. [Method] Clarify whether the combinatorial graph partitioning introduces any tunable hyperparameters (e.g., relative weighting between dissimilarity and dispersion terms) and, if so, how they are set across experiments.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the major comments point by point below, clarifying the experimental support in the full manuscript and committing to targeted revisions.

read point-by-point responses

  1. Referee: [Abstract / method description of soft pose propagation] The central claim that the diversity-aware chunks produce the reported gains rests on the soft pose propagation strategy yielding a sufficiently accurate proxy for spatial dispersion. The manuscript must supply ablations that compare the full method against (i) visual-similarity-only partitioning and (ii) random balanced chunks on the same backbones and datasets; without these controls it is impossible to attribute improvements to the spatial-dispersion term rather than to chunking in general.

    Authors: We agree that isolating the contribution of the approximated spatial dispersion term is important for validating the central claim. The current experiments focus on the complete diversity-aware partitioning pipeline, but we will add the requested ablations in the revised manuscript: direct comparisons of the full method against (i) visual-similarity-only partitioning and (ii) random balanced chunks, evaluated on the same backbones and datasets. These will be presented alongside the existing results to attribute gains specifically to the spatial dispersion component. revision: yes

  2. Referee: [Abstract / experimental section] No quantitative results, error bars, dataset statistics, or ablation tables appear in the abstract, and the provided text supplies none of the concrete numbers needed to evaluate the claimed improvements in pose estimation, depth, or reconstruction. The load-bearing experimental evidence must be presented with direct comparisons to the unpartitioned VGGT baseline.

    Authors: The full manuscript contains quantitative results, including direct comparisons to the unpartitioned VGGT baseline, along with dataset statistics in the experimental section. To address the concern about accessibility, we will revise the abstract to incorporate key quantitative improvements (with error bars where reported), dataset statistics, and explicit mention of the baseline comparisons. This will make the load-bearing evidence visible at the abstract level without altering the core claims. revision: partial

Circularity Check

0 steps flagged

No circularity detected in the proposed framework

full rationale

The manuscript describes a training-free plug-and-play inference method that partitions views via combinatorial graph partitioning on visual dissimilarity plus an approximated spatial dispersion term obtained from soft pose propagation on seed frames. No equations, derivations, or fitted parameters are shown that reduce the claimed performance gains (in pose estimation, depth prediction, or reconstruction) to quantities defined by the method itself or to self-citations whose validity depends on the present work. The approach is presented as a heuristic organization strategy whose value is asserted through external experiments rather than by algebraic identity or load-bearing self-reference; therefore the derivation chain remains self-contained against the listed circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review prevents enumeration of concrete free parameters or axioms; the method implicitly relies on standard graph-partitioning assumptions and the unstated effectiveness of visual-similarity-based spatial approximation.

pith-pipeline@v0.9.1-grok · 5741 in / 1177 out tokens · 23905 ms · 2026-07-03T16:09:08.022203+00:00 · methodology

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

48 extracted references · 5 canonical work pages · 2 internal anchors

  1. [1]

    In: CVPR (2022) 10, 12, 13, 14, 19, 21, 22, 23, 24, 26, 27, 28, 29, 30

    Azinović, D., Martin-Brualla, R., Goldman, D.B., Nießner, M., Thies, J.: Neural rgb-d surface reconstruction. In: CVPR (2022) 10, 12, 13, 14, 19, 21, 22, 23, 24, 26, 27, 28, 29, 30

    2022

  2. [2]

    In: CVPR Work- shop

    Berton, G., Masone, C.: Megaloc: One retrieval to place them all. In: CVPR Work- shop. pp. 2861–2867 (2025) 21

    2025

  3. [3]

    In: ICLR (2023) 2

    Bolya, D., Fu, C.Y., Dai, X., Zhang, P., Feichtenhofer, C., Hoffman, J.: Token merging: Your vit but faster. In: ICLR (2023) 2

    2023

  4. [4]

    In: CVPR

    Cabon, Y., Stoffl, L., Antsfeld, L., Csurka, G., Chidlovskii, B., Revaud, J., Leroy, V.: Must3r: Multi-view network for stereo 3d reconstruction. In: CVPR. pp. 1050– 1060 (2025) 12, 13

    2025

  5. [5]

    In: CVPR

    Chen, X., Li, Q., Wang, T., Xue, T., Pang, J.: Gennbv: Generalizable next-best- view policy for active 3d reconstruction. In: CVPR. pp. 16436–16445 (2024) 4

    2024

  6. [6]

    In: ICRA (1985) 4

    Connolly, C.: The determination of next best views. In: ICRA (1985) 4

    1985

  7. [7]

    In: CVPR (2017) 10, 12, 19, 27, 28, 29, 31, 32

    Dai,A.,Chang,A.X.,Savva,M.,Halber,M.,Funkhouser,T.,Nießner,M.:Scannet: Richly-annotated 3d reconstructions of indoor scenes. In: CVPR (2017) 10, 12, 19, 27, 28, 29, 31, 32

    2017

  8. [8]

    In: ICRA (2026) 2, 4, 12, 13, 25

    Deng, K., Ti, Z., Xu, J., Yang, J., Xie, J.: Vggt-long: Chunk it, loop it, align it – pushing vggt’s limits on kilometer-scale long rgb sequences. In: ICRA (2026) 2, 4, 12, 13, 25

    2026

  9. [9]

    In: ECCV

    Izquierdo, S., Civera, J.: Close, but not there: Boosting geographic distance sensi- tivity in visual place recognition. In: ECCV. pp. 240–257. Springer (2024) 23

    2024

  10. [10]

    In: CVPR (June 2024) 21

    Izquierdo,S.,Civera,J.:Optimaltransportaggregationforvisualplacerecognition. In: CVPR (June 2024) 21

    2024

  11. [11]

    arXiv preprint arXiv:2510.23928 (2025) 4

    Jha, R., Zhou, Y., Loianno, G.: Adaptive keyframe selection for scalable 3d scene reconstruction in dynamic environments. arXiv preprint arXiv:2510.23928 (2025) 4

    work page arXiv 2025

  12. [12]

    In: CVPR (2026) 2

    Kang, G., Yang, S., Nam, S., Lee, Y., Kim, J., Park, E.: Multi-view pyramid transformer: Look coarser to see broader. In: CVPR (2026) 2

    2026

  13. [13]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition

    Keetha, N., Karhade, J., Jatavallabhula, K.M., Yang, G., Scherer, S., Ramanan, D., Luiten, J.: Splatam: Splat track & map 3d gaussians for dense rgb-d slam. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition. pp. 21357–21366 (2024) 4

    2024

  14. [14]

    The Bell System Technical Journal49(2), 291–307 (1970) 4, 8, 14, 20

    Kernighan, B.W., Lin, S.: An efficient heuristic procedure for partitioning graphs. The Bell System Technical Journal49(2), 291–307 (1970) 4, 8, 14, 20

    1970

  15. [15]

    In: CVPR

    Kim, Y., Song, W., Lew, J., Hwangbo, H., Lee, J., Yoon, S.: Hess: Head sensitivity score for sparsity redistribution in vggt. In: CVPR. pp. 36509–36517 (2026) 2

    2026

  16. [16]

    ACM Transactions on Graphics36(4) (2017) 10, 12, 13, 20, 24

    Knapitsch, A., Park, J., Zhou, Q.Y., Koltun, V.: Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics36(4) (2017) 10, 12, 13, 20, 24

    2017

  17. [17]

    arXiv preprint arXiv:2511.18290 (2025) 2, 4

    Lee, J., Lee, M., Yang, S., Kang, M., Lee, S.: Swiftvggt: A scalable visual geom- etry grounded transformer for large-scale scenes. arXiv preprint arXiv:2511.18290 (2025) 2, 4

    work page arXiv 2025

  18. [18]

    In: ECCV (2024) 4

    Leroy, V., Cabon, Y., Revaud, J.: Grounding image matching in 3d with mast3r. In: ECCV (2024) 4

    2024

  19. [19]

    In: ICCV

    Li, X., Rao, T., Pan, C.: Edm: Efficient deep feature matching. In: ICCV. pp. 26198–26208 (2025) 4

    2025

  20. [20]

    In: ICCV (2023) 4 Diversity-aware View Partitioning for Scalable VGGT 17

    Lindenberger, P., Sarlin, P.E., Pollefeys, M.: LightGlue: Local Feature Matching at Light Speed. In: ICCV (2023) 4 Diversity-aware View Partitioning for Scalable VGGT 17

    2023

  21. [21]

    IEEE transactions on robotics31(5), 1147–1163 (2015) 4

    Mur-Artal, R., Montiel, J.M.M., Tardos, J.D.: Orb-slam: A versatile and accurate monocular slam system. IEEE transactions on robotics31(5), 1147–1163 (2015) 4

    2015

  22. [22]

    TMLR (2024) 4, 7, 21, 23

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. TMLR (2024) 4, 7, 21, 23

    2024

  23. [23]

    In: IROS (2019) 10, 11, 19, 26, 27

    Palazzolo, E., Behley, J., Lottes, P., Giguere, P., Stachniss, C.: Refusion: 3d re- construction in dynamic environments for rgb-d cameras exploiting residuals. In: IROS (2019) 10, 11, 19, 26, 27

    2019

  24. [24]

    In: ECCV (2024) 4

    Pan, L., Barath, D., Pollefeys, M., Schönberger, J.L.: Global Structure-from- Motion Revisited. In: ECCV (2024) 4

    2024

  25. [25]

    In: ICCV (2021) 7

    Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: ICCV (2021) 7

    2021

  26. [26]

    In: CVPR

    Sarlin, P.E., DeTone, D., Malisiewicz, T., Rabinovich, A.: Superglue: Learning feature matching with graph neural networks. In: CVPR. pp. 4938–4947 (2020) 4

    2020

  27. [27]

    In: CVPR (2016) 4, 19

    Schönberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: CVPR (2016) 4, 19

    2016

  28. [28]

    Psychometrika31(1), 1–10 (1966) 25

    Schönemann, P.H.: A generalized solution of the orthogonal procrustes problem. Psychometrika31(1), 1–10 (1966) 25

    1966

  29. [29]

    ACM Computing Surveys35(1), 64–96 (2003) 4

    Scott, W.R., Roth, G., Rivest, J.F.: View planning for automated three- dimensional object reconstruction and inspection. ACM Computing Surveys35(1), 64–96 (2003) 4

    2003

  30. [30]

    In: ICLR (2026),https://openreview.net/forum? id=asl8NJlIMe2, 3, 4, 10, 13, 19

    Shen, Y., Zhang, Z., Qu, Y., Zheng, X., Ji, J., Zhang, S., Cao, L.: FastVGGT: Fast visual geometry transformer. In: ICLR (2026),https://openreview.net/forum? id=asl8NJlIMe2, 3, 4, 10, 13, 19

    2026

  31. [31]

    In: CVPR (2013) 10, 11, 12, 14, 15, 19, 20, 22, 23, 27, 34

    Shotton, J., Glocker, B., Zach, C., Izadi, S., Criminisi, A., Fitzgibbon, A.: Scene coordinate regression forests for camera relocalization in rgb-d images. In: CVPR (2013) 10, 11, 12, 14, 15, 19, 20, 22, 23, 27, 34

    2013

  32. [32]

    In: CVPR

    Shu, Z., Lin, C., Xie, T., Yin, W., Li, B., Pu, Z., Li, W., Yao, Y., Cao, X., Guo, X., Long, X.X.: Litevggt: Boosting vanilla vggt via geometry-aware cached token merging. In: CVPR. pp. 36422–36432 (June 2026) 2, 3, 4, 5, 10, 11, 12, 13, 26, 27, 28

    2026

  33. [33]

    Splatt3R: Zero-shot Gaussian Splatting from Uncalibrated Image Pairs

    Smart, B., Zheng, C., Laina, I., Prisacariu, V.A.: Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs. arXiv preprint arXiv:2408.13912 (2024) 4

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  34. [34]

    In: IROS (Oct 2012) 10, 12, 13, 19, 21, 24

    Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for the evaluation of rgb-d slam systems. In: IROS (Oct 2012) 10, 12, 13, 19, 21, 24

    2012

  35. [35]

    In: SIGGRAPH (2022) 19, 25

    Sun, J., Chen, X., Wang, Q., Li, Z., Averbuch-Elor, H., Zhou, X., Snavely, N.: Neural 3D reconstruction in the wild. In: SIGGRAPH (2022) 19, 25

    2022

  36. [36]

    In: CVPR

    Sun, J., Shen, Z., Wang, Y., Bao, H., Zhou, X.: Loftr: Detector-free local feature matching with transformers. In: CVPR. pp. 8922–8931 (June 2021) 4

    2021

  37. [37]

    In: CVPR

    Sun, X., Zhu, Z., Lou, Z., Yang, B., Tang, J., Zhang, L., Wang, H., Zhang, J.: Avggt: Rethinking global attention for accelerating vggt. In: CVPR. pp. 251–260 (June 2026) 2, 4

    2026

  38. [38]

    Block-Sparse Global Attention for Efficient Multi-View Geometry Transformers

    Wang, C.S.B., Schmidt, C., Piekenbrinck, J., Leibe, B.: Faster vggt with block- sparse global attention. arXiv preprint arXiv:2509.07120 (2025) 2, 3, 4, 5, 10, 11, 12, 13, 26, 27, 28

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  39. [39]

    Wang, H., Agapito, L.: 3d reconstruction with spatial memory. In: 3DV. pp. 78–89. IEEE (2025) 12, 13

    2025

  40. [40]

    In: CVPR (2025) 1, 2, 4, 5, 7, 10, 11, 12, 13, 14, 15, 21, 24, 25, 26, 27, 28 18 J

    Wang, J., Chen, M., Karaev, N., Vedaldi, A., Rupprecht, C., Novotny, D.: Vggt: Visual geometry grounded transformer. In: CVPR (2025) 1, 2, 4, 5, 7, 10, 11, 12, 13, 14, 15, 21, 24, 25, 26, 27, 28 18 J. Parket al

    2025

  41. [41]

    In: CVPR (2024) 4

    Wang, J., Karaev, N., Rupprecht, C., Novotny, D.: Vggsfm: Visual geometry grounded deep structure from motion. In: CVPR (2024) 4

    2024

  42. [42]

    In: CVPR (2024) 4

    Wang, S., Leroy, V., Cabon, Y., Chidlovskii, B., Revaud, J.: Dust3r: Geometric 3d vision made easy. In: CVPR (2024) 4

    2024

  43. [43]

    In: CVPR

    Wang, W., Meiner, L., Shubham, R., De La Parra, C., Kumar, A.: Httm: Head- wise temporal token merging for faster vggt. In: CVPR. pp. 26379–26388 (June

  44. [44]

    In: ICLR (2026) 10, 12, 13

    Wang, Y., Zhou, J., Zhu, H., Chang, W., Zhou, Y., Li, Z., Chen, J., Pang, J., Shen, C., He, T.:π3: Permutation-equivariant visual geometry learning. In: ICLR (2026) 10, 12, 13

    2026

  45. [45]

    In: CVPR

    Wang, Z., Xu, D.: Flashvggt: Efficient and scalable visual geometry transformers with compressed descriptor attention. In: CVPR. pp. 21826–21835 (June 2026) 2, 4

    2026

  46. [46]

    In: NeurIPS (2022) 4

    Weinzaepfel, P., Leroy, V., Lucas, T., BRÉGIER, R., Cabon, Y., ARORA, V., Antsfeld, L., Chidlovskii, B., Csurka, G., Revaud, J.: Croco: Self-supervised pre- training for 3d vision tasks by cross-view completion. In: NeurIPS (2022) 4

    2022

  47. [47]

    In: CVPR

    Wilson, J., Almeida, M., Mahajan, S., Labrie, M., Ghaffari, M., Ghasemalizadeh, O., Sun, M., Kuo, C.H., Sen, A.: Pop-gs: Next best view in 3d-gaussian splatting with p-optimality. In: CVPR. pp. 3646–3655 (2025) 4

    2025

  48. [48]

    Yang, J., Sax, A., Liang, K.J., Henaff, M., Tang, H., Cao, A., Chai, J., Meier, F., Feiszli, M.: Fast3r: Towards 3d reconstruction of 1000+ images in one forward pass. In: CVPR (2025) 2 Diversity-aware View Partitioning for Scalable VGGT 19 Supplementary Materials Diversity-aware View Partitioning for Scalable VGGT A Additional Implementation Details A.1 ...

    work page arXiv 2025