Pith. sign in

REVIEW 2 major objections 5 minor 43 references

Frozen VGGT already encodes co-visibility in a hierarchy of layers; a tiny mixture-of-experts head extracts it better than humans on sparse indoor views.

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 · grok-4.5

2026-07-13 02:33 UTC pith:YT7OGVAL

load-bearing objection Frozen VGGT already carries a usable co-visibility signal; a tiny MoE head turns it into clear SOTA on Co-VisiON with well-calibrated pairwise scores. the 2 major comments →

arxiv 2607.09503 v1 pith:YT7OGVAL submitted 2026-07-10 cs.CV cs.AI

What VGGT Knows About Overlap: Probing Geometric Foundation Models for Co-Visibility

Filippo Ziliotto , Luciano Serafini , Lamberto Ballan , Tommaso Campari This is my paper
classification cs.CV cs.AI
keywords co-visibilitygeometric foundation modelsVGGTmixture-of-expertsmultiview geometryembodied perceptionlayer specializationvisibility graphs
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.

Co-visibility—deciding which camera views share any overlapping surface—is a make-or-break step for 3D reconstruction and robot localization when images barely overlap. This paper shows that the pretrained Visual Geometry Grounded Transformer already contains that signal as an emergent property: early layers assemble a 3D-aware scene model while late layers specialize in judging overlap, with layer 17 consistently flagging non-overlapping pairs. By freezing the backbone and training only a lightweight layer-wise mixture-of-experts head (under 7.5 million parameters) on RGB pairs, the authors obtain Co-VGGT. On the Co-VisiON benchmark it exceeds the human annotation baseline and lifts prior methods by more than 25 percent pairwise and about 10 percent multiview, while producing well-calibrated scores that can be used directly as edge weights in visibility graphs for SfM and SLAM.

Core claim

VGGT, trained without any co-visibility supervision, spontaneously develops a hierarchical internal structure in which early transformer layers build 3D-aware scene representations and late layers function as dedicated co-visibility reasoners; layer L17 in particular acts as a reliable negative anchor that routes non-co-visible pairs regardless of evaluation setting. Freezing this backbone and attaching a trainable layer-wise mixture-of-experts head that treats each layer as an expert yields Co-VGGT, which classifies co-visibility from RGB alone at near-human accuracy and with excellent calibration.

What carries the argument

Layer-wise mixture-of-experts head: each VGGT layer is treated as a specialized expert whose pair embedding is scored by a small MLP; a gating network adaptively weights the experts per image pair, exposing that co-visibility decisions concentrate in the late layers and that L17 consistently rejects non-overlapping pairs.

Load-bearing premise

The observed gating pattern and L17 negative-anchor behavior truly reflect specialized geometric reasoning inside VGGT rather than an artifact of how the lightweight head constructs pair features or is trained.

What would settle it

Retrain the same MoE head on randomly permuted or early-only layer features; if L17 ceases to dominate negative routing and accuracy collapses to chance, the claimed hierarchical specialization is not present in the backbone.

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

If this is right

  • Pairwise co-visibility scores can be inserted directly as continuous edge weights in visibility graphs for SfM and SLAM without temperature scaling or post-hoc recalibration.
  • A lightweight co-visibility head can serve as an early-failure detector that flags inconsistent geometric constraints before reconstruction drifts.
  • The same probing recipe can be reused to locate other geometric primitives inside geometry-grounded foundation models without fine-tuning the backbone.
  • In sparse-view robotic mapping the method remains reliable even when average pairwise overlap falls below 4 percent, exactly the regime where classical matchers fail.

Where Pith is reading between the lines

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

  • If late-layer specialization is a general property of geometry transformers, similar mixture-of-experts probes could extract continuous overlap ratios or occlusion maps with comparable efficiency.
  • The pairwise–multiview calibration gap suggests that a true set-level aggregator (rather than independent pair loops) would both raise multiview accuracy and restore the sharp negative-anchor routing observed in pairwise mode.
  • Attention-map signatures of failed cross-view correspondence at L17 could become a cheap diagnostic for deciding when to request a new viewpoint in active mapping.

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. The paper shows that the frozen VGGT geometric foundation model encodes co-visibility as an emergent hierarchical signal (early layers for 3D-aware scene structure, late layers for overlap reasoning, with L17 consistently acting as a negative anchor for non-co-visible pairs). It introduces Co-VGGT, which freezes VGGT and trains only a lightweight (~7.5M-parameter) layer-wise mixture-of-experts head that treats each transformer layer as an expert and adaptively gates them to classify co-visibility from RGB alone. On the Co-VisiON benchmark (Gibson and HM3D), Co-VGGT exceeds the reported human annotation baseline and prior methods by >25% pairwise IoU* and ~10% multiview IoU*, with pairwise scores well-calibrated (ECE=0.030) for direct use as visibility-graph edge weights; supporting evidence includes zero-shot cosine baselines, layer-count and aggregator ablations, difficulty-bin robustness, cross-domain transfer, and a COLMAP pair-selection proof-of-concept.

Significance. If the empirical results hold, the work is significant for geometric foundation models and embodied perception: it shows that rich, usable co-visibility priors already exist inside a frozen multiview transformer and can be distilled with a tiny head, yielding practical, calibrated edge weights for SfM/SLAM visibility graphs without post-hoc recalibration or backbone fine-tuning. Explicit strengths include the extensive ablation suite (layer ranges, pair features, zero-shot, cross-domain Gibson↔HM3D, difficulty bins, ECE/MCE/Brier, COLMAP), public code/data, and the clear separation of frozen backbone from trainable head, which makes the performance claims falsifiable and reproducible. The hierarchical-emergence interpretation, while correlational, supplies a useful organizing hypothesis that aligns with known LLM layer specialization and invites further mechanistic study.

major comments (2)
  1. [Sec. 4.4 / Fig. 6] Sec. 4.4, Fig. 6 and the accompanying attention-map discussion: the central interpretive claim that VGGT exhibits genuine task-grounded layer specialization (early layers geometric primitives, late layers dedicated co-visibility reasoners, L17 a consistent negative anchor) rests on MoE gating weights and post-hoc observations. These are correlational; the head could simply learn to ignore early layers for any discriminative signal under the chosen pair-feature construction (Eq. 2) and summarizer. Tab. 5 (early-layer collapse) and the zero-shot late-layer preference (Tab. 6) support the pattern but do not establish causality. A short causal check (layer ablation with frozen gates, or targeted activation patching) or a clearer statement that the hierarchy is an empirical observation rather than a proven internal mechanism would make the claim load-bearing-safe.
  2. [Sec. 4.2 / Tab. 2] Sec. 4.2, Sec. 5 and Tab. 2: multiview IoU* (0.74 Gibson) lags pairwise (0.85) and is explained by noisier fixed-size embeddings obtained by looping over pairs rather than a native multiview aggregator. Because the paper markets both regimes and claims multiview utility for larger view sets, this architectural gap is load-bearing for the multiview half of the contribution; either a true set-level aggregation experiment or a sharper limitation statement is needed so readers do not over-interpret the multiview numbers.
minor comments (5)
  1. [Fig. 5] Fig. 5 caption and surrounding text: the cross-view similarity matrices are described as “attention token similarity” from layer 17, yet the main text also refers to them as attention maps; a single consistent term and a brief note on how the matrices are computed would remove ambiguity.
  2. [Sec. 3] Eq. (1)–(4) and the MoE diagram (Fig. 3): the projection dimension C_proj, number of summary tokens T, and expert/gate MLP widths are free parameters never listed in the main text (only total head size ~7.5 M is given). A short hyper-parameter table or appendix listing would aid exact reproduction.
  3. [Tab. 3] Table 3 difficulty bins: the edge-level and graph-level definitions are clear, but the number of pairs/scenes falling into each bin is not reported; adding those counts would let readers judge whether the hard-bin gains are statistically stable.
  4. [Supp. A.5] Supplementary Sec. A.5 COLMAP experiment: the 20-scene subset and top-30 % budget are useful, yet the absolute numbers of registered images and points for the exhaustive baseline are modest; a note on scene selection criteria would clarify generalizability.
  5. Minor typographical inconsistencies appear (e.g., “co-visibility” vs “co-visibile”, occasional missing spaces around citations). A final proof-reading pass would polish presentation.

Circularity Check

0 steps flagged

No significant circularity: frozen external VGGT backbone plus supervised head on external Co-VisiON labels yields empirical results, not a self-referential derivation.

full rationale

The paper's central claims rest on an external pretrained model (VGGT [34], frozen, never trained on co-visibility) whose features are probed by a lightweight trainable MoE head (<7.5 M parameters) supervised on the external Co-VisiON benchmark [8] with scene-disjoint splits. Zero-shot cosine similarity on frozen layer embeddings already produces non-trivial signal (Tables 1, 6), layer-count ablations collapse when early layers alone are used (Table 5), and pairwise ECE=0.030 is measured on held-out data. The hierarchical/L17-negative-anchor observations (Fig. 6, Sec. 4.4) are post-hoc correlational analyses of the trained gate, not definitional of the target labels or of the performance numbers. No parameter is fitted to a quantity that is then re-presented as a prediction of the same quantity; no uniqueness theorem or ansatz is imported from overlapping authors to force the result; and the method is evaluated against independent baselines (Covis, DUSt3R, VLMs, human annotation). The derivation chain is therefore ordinary empirical probing plus supervised classification, fully self-contained against external data and models.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The central claims rest on the pretrained VGGT backbone (taken as a black-box geometric foundation model), the Co-VisiON definition of binary co-visibility from surface overlap, standard transformer probing assumptions, and a small set of architectural and optimization choices for the MoE head. No new physical entities are postulated; the 'negative anchor' is an observed routing pattern, not an invented mechanism.

free parameters (4)
  • MoE expert/gate MLP widths and projection dimension C_proj / T_summary
    Architectural sizes of the trainable head (~7.5 M parameters total) chosen by the authors; not derived from first principles.
  • Learning rate 1e-4, weight decay 1e-4, batch size 32, 50 epochs
    Standard optimization hyper-parameters selected for best validation AUC; affect absolute numbers but not the qualitative hierarchy claim.
  • Layer range used for MoE experts (default all 24, ablated to last 12 etc.)
    Choice of which VGGT layers feed the head; ablations show last layers dominate, but the exact cut-off is empirical.
  • IoU-optimal threshold τ* and ECE binning
    Post-training operating-point and calibration choices used to report headline metrics.
axioms (4)
  • domain assumption VGGT was pretrained without any co-visibility or overlap supervision, so any co-visibility signal in its features is emergent.
    Stated in abstract and introduction; required for the 'emergent behavior' claim.
  • domain assumption Co-visibility is well-defined as non-zero shared visible surface area between two views, and Co-VisiON labels correctly capture it.
    Problem setup (Sec. 3) and evaluation protocol; all metrics rest on these binary labels.
  • ad hoc to paper Layer-wise gating weights of a trained MoE head reflect the intrinsic specialization of the frozen backbone layers.
    Core interpretive move in Sec. 4.4; correlational, not causally proven by interventions.
  • domain assumption Standard transformer residual stream and attention mechanisms allow hierarchical geometric abstraction analogous to LLM layer specialization.
    Invoked when claiming the hierarchy 'mirrors that of large language models'.
invented entities (2)
  • Co-VGGT (layer-wise MoE head on frozen VGGT) independent evidence
    purpose: Lightweight trainable classifier that turns multi-layer VGGT embeddings into calibrated co-visibility probabilities.
    New architecture introduced by the paper; independent evidence is the reported Co-VisiON numbers and ablations.
  • L17 negative anchor no independent evidence
    purpose: Observed consistent routing of non-co-visible pairs to layer 17, used as evidence of layer specialization.
    Descriptive label for an empirical gating peak; not an independent physical entity, but treated as a stable property of this backbone.

pith-pipeline@v1.1.0-grok45 · 23070 in / 3266 out tokens · 39547 ms · 2026-07-13T02:33:10.811305+00:00 · methodology

0 comments
read the original abstract

A fundamental challenge in 3D reconstruction and robotic localization is co-visibility: determining which image pairs share overlapping visible surfaces, particularly in scenarios with minimal overlap. We demonstrate that VGGT implicitly encodes co-visibility as an emergent behavior: without any supervision for this task, its internal representations exhibit a clear hierarchical structure mirroring that of large language models, i.e. early layers build a 3D-aware scene representation, while late layers act as dedicated co-visibility reasoners. In particular, we identify layer L17 as a negative anchor that consistently routes non-co-visible pairs for this backbone, regardless of the evaluation setting, providing task-grounded evidence of layer specialization in a geometry-grounded foundation model. Building on this, we introduce Co-VGGT, which freezes VGGT and trains only a lightweight layer-wise mixture-of-experts head (less than 7.5M parameters) to classify co-visibility from RGB alone, treating each layer as a specialized expert whose geometric abstraction is adaptively weighted per input pair. On the Co-VisiON benchmark, Co-VGGT surpasses the human annotation baseline and improves over prior work by more than 25% pairwise and 10% multiview. Pairwise predictions are well-calibrated (ECE=0.030), enabling direct use as edge weights in visibility graphs for downstream SfM and SLAM pipelines without post-hoc correction. Code and data are available.

Figures

Figures reproduced from arXiv: 2607.09503 by Filippo Ziliotto, Lamberto Ballan, Luciano Serafini, Tommaso Campari.

Figure 1
Figure 1. Figure 1: Co-visibility Task. We study co-visibility prediction: given multiple RGB views of a scene, the goal is to determine which image pairs share overlapping vis￾ible regions. Our approach probes geometric consistency in a frozen VGGT founda￾tion model, extracting layer-wise view embeddings and combining them through a lightweight mixture-of-experts head. Without modifying the backbone, the model ag￾gregates in… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the Co-VGGT method. Input RGB views are processed by the frozen VGGT backbone to extract layer-wise features. These features are projected, summarized, and formed into per-view embeddings. In both pairwise and multiview modes, these embeddings are used to construct pair features, which are then fed into a trainable Mixture-of-Experts (MoE) head to predict co-visibility probabilities, enabling t… view at source ↗
Figure 3
Figure 3. Figure 3: Co-visibility Estimation Head. Given a set of sparse RGB views (left), our method leverages the frozen Visual Geometry Grounded Transformer (VGGT) to extract layer-wise view embeddings. A lightweight, trainable Mixture-of-Experts (MoE) head (center) adaptively aggregates these multi-scale features to predict pair￾wise co-visibility probabilities. The resulting predictions form a dense scene-level vis￾ibili… view at source ↗
Figure 4
Figure 4. Figure 4: Co-visibility Scene-Graph Examples (Multiview). [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Cross-view Similarity Matrix Examples. Examples of Co-VGGT atten￾tion token similarity over pair of co-visibile images. (left) input RGB image, (right) the attention output mask. Features are extracted from signal belonging to layer 17 of the Pairwise evaluation [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: MoE Gating weights. Average α parameter per-layer vs. layer id. on the multiview (left) and pairwise (right) task. We observe that specific layers are decisive for the final co-visibility analysis. is driven by sparsely populated mid-confidence bins, since over 90% of predictions fall in [0, 0.1) or [0.9, 1]. Furthermore, we added a proof-of-concept COLMAP experiment against different methods to show this … view at source ↗
Figure 7
Figure 7. Figure 7: Calibration Measures. Calibration plot (left), score distribution (center) and summary metrics (right) on Gibson validation for pairwise and multiview tasks. sized in [34]: early layers carry no discriminative signal for co-visibility, while late layers encode the high-level judgment of accepting or rejecting shared surface support, with positive and negative pairs peaking at distinct layers (L15 and L17, … 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

43 extracted references · 2 canonical work pages

  1. [1]

    arXiv preprint arXiv:2509.15880 (2025).https: //doi.org/10.48550/arXiv.2509.15880

    An, V.D., Vu, M.N., Reid, I.D.: Improving robotic manipulation with efficient geometry-aware vision encoder. arXiv preprint arXiv:2509.15880 (2025).https: //doi.org/10.48550/arXiv.2509.15880

    doi:10.48550/arxiv.2509.15880 2025

  2. [2]

    In: CVPR (2016)

    Arandjelovic, R., Gronat, P., Torii, A., Pajdla, T., Sivic, J.: Netvlad: Cnn archi- tecture for weakly supervised place recognition. In: CVPR (2016)

    2016

  3. [3]

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., Zhong, H., Zhu, Y., Yang, M., Li, Z., Wan, J., Wang, P., Ding, W., Fu, Z., Xu, Y., Ye, J., Zhang, X., Xie, T., Cheng, Z., Zhang, H., Yang, Z., Xu, H., Lin, J.: Qwen2.5-vl technical report (2025),https://arxiv.org/abs/2502.13923

    Pith/arXiv arXiv 2025

  4. [4]

    arXiv preprint arXiv:2202.12197 (2022)

    Bavle, H., Sanchez-Lopez, J.L., Civera, J., Voos, H.: Situational graphs for robot navigation in structured indoor environments. arXiv preprint arXiv:2202.12197 (2022)

    Pith/arXiv arXiv 2022

  5. [5]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Cabon, Y., Stoffl, L., Antsfeld, L., Csurka, G., Chidlovskii, B., Revaud, J., Leroy, V.: Must3r: Multi-view network for stereo 3d reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1050–1060 (2025).https://doi.org/10.1109/CVPR52734.2025.00106

    doi:10.1109/cvpr52734.2025.00106 2025

  6. [6]

    In: T-RO

    Campos, C., Elvira, R., Rodríguez, J.J.G., Montiel, J., Tardós, J.D.: Orb-slam3: An accurate open-source library for visual, visual–inertial, and multimap slam. In: T-RO. vol. 37, pp. 1874–1890 (2021)

    2021

  7. [7]

    In: ICCV (2021)

    Chen, A., Xu, Z., Zhao, J., Yu, J., Su, H., Zhang, J., Yu, J.: Mvsnerf: Fast gener- alizable radiance field reconstruction from multi-view stereo. In: ICCV (2021)

    2021

  8. [8]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops

    Chen, C., Dang, N., Zhang, J., Sun, W., Zheng, P., He, X., Ye, Y., Zhang, J., Srini- vas, T., Feng, C.: Co-vision: Co-visibility reasoning on sparse image sets of indoor scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 4861–4871 (October 2025)

    2025

  9. [9]

    In: NeurIPS (2024)

    Cheng, A.C., Yin, H., Fu, Y., Guo, Q., Yang, R., Kautz, J., Wang, X., Liu, S.: Spatial-rgpt: Grounded spatial reasoning in vision-language models. In: NeurIPS (2024)

    2024

  10. [10]

    In: Proceedings of the 2017 conference on empirical methods in natural language processing

    Conneau, A., Kiela, D., Schwenk, H., Barrault, L., Bordes, A.: Supervised learn- ing of universal sentence representations from natural language inference data. In: Proceedings of the 2017 conference on empirical methods in natural language processing. pp. 670–680 (2017)

    2017

  11. [11]

    arXiv preprint arXiv:2010.11929 (2020)

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    Pith/arXiv arXiv 2010

  12. [12]

    Geva, M., Katz, U., Ben-Arie, A., Berant, J.: What’s in your head? emergent behaviourinmulti-tasktransformermodels.In:Proceedingsofthe2021Conference on Empirical Methods in Natural Language Processing. pp. 8201–8215 (2021)

    2021

  13. [13]

    In: Pro- ceedings of the 22nd ACM SIGKDD international conference on Knowledge dis- covery and data mining

    Grover, A., Leskovec, J.: node2vec: Scalable feature learning for networks. In: Pro- ceedings of the 22nd ACM SIGKDD international conference on Knowledge dis- covery and data mining. pp. 855–864 (2016)

    2016

  14. [14]

    He,K.,Zhang,X.,Ren,S.,Sun,J.:Deepresiduallearningforimagerecognition.In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770–778 (2016)

    2016

  15. [15]

    arXiv preprint arXiv:2201.13360 (2022) What VGGT Knows About Overlap 17

    Hughes, N., Chang, Y., Carlone, L.: Hydra: A real-time spatial perception system for 3d scene graph construction and optimization. arXiv preprint arXiv:2201.13360 (2022) What VGGT Knows About Overlap 17

    Pith/arXiv arXiv 2022

  16. [16]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Jiang, H., Karpur, A., Cao, B., Huang, Q.: Omniglue: Generalizable feature match- ing with foundation model guidance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19865–19875 (2024). https://doi.org/10.1109/CVPR52733.2024.01878

    doi:10.1109/cvpr52733.2024.01878 2024

  17. [17]

    arXiv preprint arXiv:2506.09883 (2025).https://doi.org/10.48550/arXiv.2506.09883

    Lee, S., Choi, J., Kang, I., Kim, J., Park, J., Shim, H.: 3d-aware vision-language models fine-tuning with geometric distillation. arXiv preprint arXiv:2506.09883 (2025).https://doi.org/10.48550/arXiv.2506.09883

    work page doi:10.48550/arxiv.2506.09883 2025

  18. [18]

    Scholarpedia7, 10491 (05 2012)

    Lindeberg, T.: Scale invariant feature transform. Scholarpedia7, 10491 (05 2012). https://doi.org/10.4249/scholarpedia.10491

    work page doi:10.4249/scholarpedia.10491 2012

  19. [19]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Lindenberger, P., Sarlin, P.E., Pollefeys, M.: Lightglue: Local feature matching at light speed. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 17581–17592 (2023).https://doi.org/10.1109/ ICCV51070.2023.01616

    arXiv 2023

  20. [20]

    Advances in neural information processing systems35, 17359–17372 (2022)

    Meng, K., Bau, D., Andonian, A., Belinkov, Y.: Locating and editing factual associ- ations in gpt. Advances in neural information processing systems35, 17359–17372 (2022)

    2022

  21. [21]

    In: Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)

    Mou, L., Men, R., Li, G., Xu, Y., Zhang, L., Yan, R., Jin, Z.: Natural language inference by tree-based convolution and heuristic matching. In: Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). pp. 130–136 (2016)

    2016

  22. [22]

    arXiv preprint arXiv:2303.08774 (2023)

    OpenAI: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

    Pith/arXiv arXiv 2023

  23. [23]

    arXiv preprint arXiv:2109.08238 (2021)

    Ramakrishnan, S.K., Gokaslan, A., Wijmans, E., Maksymets, O., et al.: Habitat- matterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai. arXiv preprint arXiv:2109.08238 (2021)

    Pith/arXiv arXiv 2021

  24. [24]

    IEEE International Con- ference on Robotics and Automation (ICRA) (2020)

    Rosinol, A., Abate, M., Chang, Y., Carlone, L.: Kimera: an open-source library for real-time metric-semantic localization and mapping. IEEE International Con- ference on Robotics and Automation (ICRA) (2020)

    2020

  25. [25]

    In: CVPR (2020)

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

    2020

  26. [26]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Schonberger, J.L., Frahm, J.M.: Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4104–4113 (2016)

    2016

  27. [27]

    arXiv preprint arXiv:1409.1556 (2014)

    Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)

    Pith/arXiv arXiv 2014

  28. [28]

    arXiv preprint arXiv:2510.24907 (2025)

    Stary, M., Gaubil, J., Tewari, A., Sitzmann, V.: Understanding multi-view trans- formers. arXiv preprint arXiv:2510.24907 (2025)

    arXiv 2025

  29. [29]

    In: ICCV (2021)

    Sucar, E., Liu, S., Ortiz, J., Davison, A.J.: imap: Implicit mapping and positioning in real-time. In: ICCV (2021)

    2021

  30. [30]

    arXiv preprint arXiv:2412.06974 (2024)

    Tang,Z.,Fan,Y.,Wang,D.,Xu,H.,Ranjan,R.,Schwing,A.,Yan,Z.:Mv-dust3r+: Single-stage scene reconstruction from sparse views in 2 seconds. arXiv preprint arXiv:2412.06974 (2024)

    Pith/arXiv arXiv 2024

  31. [31]

    arXiv preprint arXiv:2312.11805 (2023)

    Team, G., Anil, R., Borgeaud, S., Alayrac, J.B., et al.: Gemini: A family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)

    Pith/arXiv arXiv 2023

  32. [32]

    In: NeurIPS (2021)

    Teed, Z., Deng, J.: Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. In: NeurIPS (2021)

    2021

  33. [33]

    In: Pro- ceedingsofthe57thannualmeetingoftheassociationforcomputationallinguistics

    Tenney, I., Das, D., Pavlick, E.: Bert rediscovers the classical nlp pipeline. In: Pro- ceedingsofthe57thannualmeetingoftheassociationforcomputationallinguistics. pp. 4593–4601 (2019)

    2019

  34. [34]

    In: IEEE/CVF Conference on Computer What VGGT Knows About Overlap 1 Vision and Pattern Recognition (CVPR) 2025, Nashville, TN, USA, June 11-15,

    Wang, J., Chen, M., Karaev, N., Vedaldi, A., Rupprecht, C., Novotný, D.: Vggt: Visual geometry grounded transformer. In: IEEE/CVF Conference on Computer What VGGT Knows About Overlap 1 Vision and Pattern Recognition (CVPR) 2025, Nashville, TN, USA, June 11-15,

    2025

  35. [35]

    5294–5306

    pp. 5294–5306. Computer Vision Foundation / IEEE (2025).https://doi. org/10.1109/CVPR52734.2025.00499

    doi:10.1109/cvpr52734.2025.00499 2025

  36. [36]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Wang, S., Leroy, V., Cabon, Y., Chidlovskii, B., Revaud, J.: Dust3r: Geometric 3d vision made easy. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 20697–20709 (2024).https://doi. org/10.1109/CVPR52733.2024.01956

    doi:10.1109/cvpr52733.2024.01956 2024

  37. [37]

    In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR)

    Wang, Y., He, X., Peng, S., Tan, D., Zhou, X.: Efficient loftr: Semi-dense local feature matching with sparse-like speed. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR). pp. 21666–21675 (2024).https://doi.org/10.1109/CVPR52733.2024.02047

    doi:10.1109/cvpr52733.2024.02047 2024

  38. [38]

    Weinzaepfel, P., Lucas, T., Leroy, V., Cabon, Y., Arora, V., Brégier, R., Csurka, G., Antsfeld,L., Chidlovskii,B., Revaud, J.: Croco v2: Improved cross-view completion pre-trainingforstereomatchingandopticalflow.In:ProceedingsoftheIEEE/CVF International Conference on Computer Vision. pp. 17969–17980 (2023)

    2023

  39. [39]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Xia, F., Zamir, A.R., He, Z., Sax, A., Malik, J., Savarese, S.: Gibson env: Real- world perception for embodied agents. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 9068–9079 (2018)

    2018

  40. [40]

    In: CVPR (2023)

    Xu, Z., Fan, Y., Bao, J., Zhang, D., Li, H., Zhang, D.: Mvfusion: A multi-view diffusion model for 3d reconstruction. In: CVPR (2023)

    2023

  41. [41]

    arXiv preprint arXiv:2208.10431 (2022)

    Xue, M., Huang, Q., Zhang, H., Cheng, L., Song, J., Wu, M., Song, M.: Protop- former: Concentrating on prototypical parts in vision transformers for interpretable image recognition. arXiv preprint arXiv:2208.10431 (2022)

    arXiv 2022

  42. [42]

    In: CVPR (2021)

    Zhang, T., Usenko, V., Engel, J., Cremers, D.: Bad-slam: Bundle adjusted direct rgb-d slam. In: CVPR (2021)

    2021

  43. [43]

    disrupting

    Zhu, Z., Peng, S., Larsson, V., Lin, C.H., Bao, H., Dai, A., Nießner, M., Zhou, X.: Nice-slam: Neural implicit scalable encoding for slam. In: CVPR (2022) What VGGT Knows About Overlap: Probing Geometric Foundation Models for Co-Visibility Filippo Ziliotto1,2, Luciano Serafini2, Lamberto Ballan1, and Tommaso Campari2 1 University of Padova 2 Fondazione Br...

    2022