Pith. sign in

REVIEW 4 major objections 4 minor 33 references

MuViSeg: Multi-View Segment Correspondences from Dense Geometry Priors

T0 review · 4 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

Pith's one-line read Joint multi-view attention over image segments recovers correspondences that one-pair-at-a-time matchers miss.

desk verdict Clean regime-split result and a fair head-to-head ablation, but the transitive-correspondence mechanism is asserted, not demonstrated. read the letter →

arxiv 2607.17938 v1 pith:DPHZ7DJU submitted 2026-07-20 cs.CV

classification cs.CV MSC 68T45
keywords segmentmatchingmulti-viewcorrespondence3Dfoundationmodelsinstancesegmentationtopologicalnavigationzero-shottransitivewide-baseline
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that object-level image matching should be scored jointly across several views rather than one pair at a time. It builds three learned matching heads on frozen 3D foundation features; the central one concatenates segment descriptors from N views into a single sequence with shared self-attention, which lets an intermediate view bridge two distant views. Under zero-shot evaluation, the pairwise attention head wins at narrow baselines and outdoors, while the joint multi-view head wins at the widest baselines and in closed-loop navigation, where it raises success from 50% to 70%. A sympathetic reader would care because anything that must know 'same object, different view' — robot navigation, scene graphs, object-level mapping — depends on exactly this correspondence, and the paper offers a regime-dependent rule for which matcher to use.

What carries the argument

The load-bearing object is the joint multi-view attention head. Per-segment descriptors are obtained by masking and pooling frozen patch features from a multi-view foundation model; a DPT-style fusion exposes multi-scale spatial detail before pooling. The descriptors from all N views are concatenated into one token sequence with learnable per-view position embeddings, and shared self-attention lets any segment attend to any segment in any view. A DoubleSoftmax scorer then reads off any requested pair. This is what makes transitive correspondences possible: an intermediate view can connect two views that share almost no segments.

What would settle it

Run the closed-loop navigation comparison on the full HM3D val split with the joint model at N=3 or more. If the success-rate advantage over the pairwise head shrinks to within chance, or if a segment-triplet consistency check shows no agreement between bridge-induced and direct matches, the transitive multi-view mechanism is not carrying the reported gain.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a joint multi-view matcher — shared self-attention over segments drawn from all N views with per-view position embeddings, followed by DoubleSoftmax scoring — recovers transitive correspondences that strictly pairwise matchers cannot reach. Trained with N=4 on ScanNet++, it generalises to N=2, 6, and 8 without retraining or collapse. In stratified zero-shot tests, the joint head leads in the 90–180 degree bins where two views share few segments; a LightGlue-style pairwise head leads at narrow baselines and on outdoor data. When dropped into a topological navigation pipeline without retraining, the joint head raises success rate from 50%

Load-bearing premise

The central claim rests on treating the 10-episode HM3D minival comparison as a reliable signal, and on attributing a +20pp success gain measured at N=2 to a multi-view transitive mechanism that requires at least three views to act.

Editorial extensions

If this is right

  • The best segment matcher is regime-dependent, so downstream systems should pair a pairwise head for small viewpoint change with a joint multi-view head for wide baselines.
  • A joint multi-view head trained once at N=4 can be evaluated at other tuple sizes, which is useful when the number of available views varies at runtime.
  • In closed-loop navigation, the joint head improves the chance of reaching the goal object (success rate and soft SPL) at the cost of longer, less direct paths.
  • Learned heads gain where training data is dense but can trail a parameter-free baseline at extreme rotations; the parameter-free fallback remains useful at 135–180 degrees.
  • Multi-scale fusion before pooling adds a few AUPRC points over pooling only the last layer, on the training-domain validation.

Reading between the lines

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

  • The navigation success gain is reported with the joint head running at N=2, the query plus one map frame; at N=2 there is no intermediate view to mediate transitivity, so the +20pp cannot be attributed to the multi-view mechanism without a test at N≥3.
  • The flat-in-N result on indoor scenes suggests the transitive mechanism matters most when per-view overlap is low; a natural extension is to make the number of jointly attended views adaptive to scene overlap or motion cues.
  • A direct transitivity test on segment triplets — checking whether match(A,C) via B agrees with the direct match — would quantify how much of the wide-baseline gain comes from bridge views rather than from having more context.
  • The regime split invites a router that switches heads based on deployment-time cues such as frame rate, estimated motion, or scene type; the paper names this as a future direction.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes three segment-level matching heads built on frozen dense 3D foundation model features: a LightGlue-style pairwise head on MASt3R descriptors (LGv2), a DPT-fused VGGT pairwise head, and a joint multi-view attention head that scores segments from N views together. Under a stratified zero-shot protocol on Replica and Virtual KITTI 2, the authors report that LGv2 wins narrow baselines and outdoor transfer, while the joint head wins the widest angular bins and improves closed-loop HM3D Instance Image Navigation success rate from 50% to 70% and SPL from 45.7 to 59.1. The central claim is that joint multi-view attention recovers transitive correspondences that strictly pairwise matchers cannot reach.

Significance. If the claims hold, the paper makes a useful contribution: a systematic, controlled comparison of segment-matching heads, a zero-shot evaluation protocol with angular stratification, and a multi-view head that extrapolates from N=4 training to N=2,6,8 without retraining. The strongest comparisons are the LGv2-vs-Sinkhorn pair on a frozen MASt3R backbone and the Joint-vs-DPT pair on a frozen VGGT backbone, which isolate the head in those specific configurations. The regime-dependent conclusion (pairwise for narrow baselines, joint for wide baselines) is practically actionable. However, the evidence for the central transitive-correspondence mechanism is currently indirect, and several key comparisons lack uncertainty quantification.

major comments (4)
  1. [§4.6, Table 3; Abstract; §1 Contribution (i)] The headline closed-loop claim is that the joint multi-view matcher 'raises success rate by 20 points' on HM3D. But the Joint model is deployed at N=2 (query plus one map frame), where transitive reasoning across an intermediate view is impossible. The +20pp gain is a 7/10 vs 5/10 difference on ten episodes, which is consistent with chance; Limitation (ii) acknowledges the small sample. Please either measure the closed-loop setting with N≥3, explicitly disentangle the joint-attention-over-two-views effect from transitivity, or remove the transitive-correspondence attribution for this result.
  2. [§4.2–§4.3, Tables 1–2, footnote 1] The open-loop evidence for the multi-view advantage rests on the gap between SegVGGT-DPT and SegVGGT-DPT Joint at the 135–180° bin: +4.3 AUPRC on Replica and +3.6 on VKITTI2. Footnote 1 states that seeds are not fixed across the two VGGT rows and instructs the reader to compare only angular trends, not sub-point differences. The numbers cited as the multi-view contribution are exactly sub-point differences. Report multiple seeds with means and standard deviations, or fix seeds, and provide confidence intervals for the wide-baseline bins.
  3. [§4.4] The N-sweep varies N only for the Joint model; it does not test whether an intermediate view resolves correspondences that are unrecoverable from the two distant views alone. A direct test would compare (i) matching A–C with the pairwise head, (ii) matching A–C with the Joint head at N=2, and (iii) matching A–B–C with the Joint head at N=3, measuring recall on A–C ground-truth pairs. The comparison should also include pairwise chaining (composing A–B and B–C matches) to show the advantage is not merely the use of more views. Without such a test, the Abstract and Contribution (i) claim that joint attention recovers correspondences 'strictly pairwise matchers cannot reach' remains an architectural assertion, not a measured result.
  4. [§3.3, §4.2, §4.6] The LGv2-vs-Sinkhorn comparison does not hold the input representation fixed. SegMASt3R (Sinkhorn) pools frozen 768-dim MASt3R patch features and learns a segment-feature MLP to 24-dim; SegMASt3R+LGv2 uses MASt3R's pre-trained 24-dim local-feature head as input to a two-layer MLP and an attention head. The +4.85 (Replica) and +25.87 (VKITTI2) AUPRC gains therefore confound the matching head with the choice of backbone features, and the statement in §4.6 that 'the only difference is the cross-segment attention head' is inaccurate. Please add an LGv2 variant that consumes the same 768-dim pooled features (with an equivalent projector), or a Sinkhorn baseline on the 24-dim local features, to isolate the head contribution.
minor comments (4)
  1. [Abstract] Typo: 'matchng' should be 'matching'. Also 'raises SPL by over 13' would be more precise as '+13.44 SPL points'.
  2. [Table 1 and Figure 2] Formatting: 'T able 1' in §4.2 has an extra space. Figure 2's caption calls the crossover 'the central observation,' but there is no statistical analysis supporting the crossover; consider adding error bars or a significance test.
  3. [Footnote 1] The seed caveat is important for interpreting Tables 1–2 and should appear in the main text or in the experimental setup section, not only as a footnote.
  4. [§3.3] The justification for N-extrapolation — 'position embeddings are sparse (only indices 0–3 are exercised during training)' — is unclear. Sparse supervision of the embeddings does not by itself explain generalization to N=6,8; please clarify what property of the learned embeddings or attention enables the extrapolation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the head comparisons are measured on frozen backbones with zero-shot evaluation; the N=2 deployment gap is an experimental limitation, not a circular derivation.

full rationale

The paper's derivation chain is self-contained with respect to the claims it actually tests. All learned heads are trained only on ScanNet++ and evaluated zero-shot on Replica, Virtual KITTI 2, and HM3D; no target-benchmark data enters training. The LGv2-vs-Sinkhorn comparison holds the MASt3R backbone frozen and changes only the head; the Joint-vs-DPT comparison holds the VGGT-plus-DPT feature extraction fixed and changes only the attention/scoring stage. The closed-loop comparison replaces only the online matcher in RoboHop using a released teach-run map, so the reported delta is attributable to the matcher rather than to a fitted pipeline. The temperature clamp tau<=1.0 is a hyperparameter selected through an error-budget analysis, not a parameter fitted to the evaluation metric, so it does not constitute a fitted input presented as a prediction. The manuscript itself flags its main evidential weaknesses: 10 HM3D minival episodes (Limitations ii) and unfixed seeds between the two VGGT rows (footnote 1). The closed-loop deployment at N=2 means the 'transitive correspondences' mechanism is not directly tested in that experiment, and the wide-baseline open-loop gains lack quantified uncertainty; these are experimental gaps that weaken causal attribution, but they do not make any result equivalent to its input by construction. No load-bearing self-citation, imported uniqueness theorem, or ansatz smuggled in via citation appears. Accordingly, the circularity score is 0.

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

The central empirical claims rest on frozen foundation-model features (MASt3R, VGGT) and standard supervised training; no new physical entities are introduced. The design choices that materially affect the reported numbers are the tau clamp, the N=4 training with sparse per-view position embeddings, and the fixed DPT layer set. The architecture's feasibility hinges on the asserted pair-independence of VGGT features.

free parameters (4)
  • DoubleSoftmax temperature tau (VGGT heads) = learned; clamped to tau <= 1.0
    Learned scoring temperature materially changes the error budget; the clamp in §4.5 (cuts false-dustbin errors 36%) is applied to the VGGT rows in Tables 1-2. If the clamp was selected after inspecting test behavior, the VGGT results are partly fit.
  • Matchability loss weight lambda = 0.3
    Hand-chosen balance between match NLL and matchability BCE (§3.3); the §4.5 ablation couples it with the tau clamp.
  • Per-view position embeddings e_v (Joint model) = indices 0-3 trained (N=4)
    The N-generalization claim (§4.4) rests on these embeddings being sparse and view-relative; indices 4-7 are never supervised, so the model's behavior at N=6,8 depends on an unverified generalization assumption.
  • DPT fusion layer selection = {5,11,17,23}
    Standard DPT layer set taken from [20]; the +2.8 AUPRC DPT gain (0.858 vs 0.886 on ScanNet++ val) is measured with this fixed choice, not swept.
assumptions (4)
  • domain assumption VGGT Aggregator features are pair-independent: per-view features do not change when other views are added to or removed from the same forward pass.
    Stated in §3.2 and load-bearing for the whole joint-head design and for attributing gains to the head rather than to changed backbone inputs. If false, per-view precomputation and the head/backbone attribution collapse.
  • domain assumption MASt3R patch features are pair-dependent, so a multi-view head is architecturally impossible on this backbone.
    §3.2 and §5 use this to justify assigning the joint head to VGGT; it also means the backbone and head axes are naturally confounded in the Joint-vs-LGv2 comparison.
  • domain assumption SAM instance masks are consistent enough across viewpoints that masked average pooling of patch features yields comparable segment descriptors, and instance-ID overlap is valid ground truth.
    §3.1 and §4.1. If SAM over- or under-segments differently per view, AUPRC measures mask-level consistency rather than object identity, and the transitive-correspondence interpretation weakens.
  • domain assumption ScanNet++ co-visible pairs have a moderate-viewpoint distribution, and the LGv2 reversal at 135-180 degrees is caused by internalized training bias.
    §4.7 interpretation: plausible but not directly demonstrated; no distributional analysis of ScanNet++ viewpoints is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MuViSeg: Multi-View Segment Correspondences from Dense Geometry Priors." pith.science (2026). https://pith.science/paper/DPHZ7DJU

@misc{pith2026260717938,
  author       = {Pith},
  title        = {Pith review of: MuViSeg: Multi-View Segment Correspondences from Dense Geometry Priors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DPHZ7DJU}},
  note         = {Machine review of arXiv:2607.17938}
}
read the original abstract

Classical image correspondence is solved at the level of sparse keypoints or dense pixels, but the systems that consume these matches - object-level mapping, topological navigation, scene-graph maintenance - reason about whole objects. Recent work narrows this gap by matchng directly at the level of instance segments: a class-agnostic segmenter partitions each image, and per-segment descriptors are obtained by pooling features from large 3D foundation models over the masks. We build on this segment-level matching paradigm and propose three learned matching heads: a LightGlue-style attention head with DoubleSoftmax scoring on frozen MASt3R descriptors; a DPT-style multi-scale fusion module that exposes layered spatial detail from the VGGT foundation model before pooling; and - as our main contribution - a multi-view extension that performs joint self-attention over segments drawn from several views at once, recovering transitive correspondences that strictly pairwise matchers cannot reach. Under a stratified zero-shot protocol on Replica and Virtual KITTI 2 with controlled viewpoint baselines from 0 deg to 180 deg, the LightGlue-style head improves over a parameter-free Sinkhorn matcher on the same MASt3R backbone by +4.85 AUPRC on Replica and +25.9 AUPRC on Virtual KITTI 2. Dropped into the RoboHop topological navigation pipeline on the Habitat-Matterport 3D (HM3D) Instance Image Navigation benchmark without retraining, our multi-view variant raises success rate from 50% to 70%, and our LightGlue-style head raises SPL from 45.7 to 59.1.

Figures

Figures reproduced from arXiv: 2607.17938 by the authors.

Figure 1
Figure 1. Architecture of the three matchers (grey: frozen, teal: trainable). Top: shared template — frozen backbone → per-patch features F, mask-selected and mask-pooled to one descriptor zk per segment. The heads differ only downstream. (a) LGv2 on MASt3R: pairwise self-/cross-attention and an FFN, then a DoubleSoftmax scorer with a matchability head. (b) VGGT+DPT: a DPT fusion of Aggregator layers {5, 11, 17, 23} before po… view at source ↗
Figure 2
Figure 2. AUPRC per relative-rotation bin on Replica (left) and Virtual KITTI 2 (right). On Replica, LGv2 leads at narrow baselines but is overtaken by Joint (and even the Sinkhorn baseline) past 90◦ ; on VKITTI2 LGv2 dominates every bin. The crossover is the central observation: the best head depends on the viewpoint regime [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Effect of the number of jointly attended views N on SegVGGT-DPT Joint (trained at N=4, evaluated at N ∈ {2, 4, 6, 8} without retraining). Replica (left) is flat in N (dense overlap); on VKITTI2 (right) extra views help only at the widest baseline (71.1→78.2 AUPRC at 135–180◦ ) and saturate by N=6. dominates the error budget and the model discards true correspondences. Clamp￾ing τ ≤ 1.0 (with matchability weight λ=0.… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Temperature-clamp ablation on SegVGGT-DPT: per-query error rate by failure type, before (left) and after (right) clamping τ ≤ 1.0 (λ=0.3). Clamping cuts the dominant false-dustbin errors by 36% relative, shifting the residual budget to the harder wrong-match/false-matc…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 6 linked inside Pith

  1. [1]

    arXiv preprint arXiv:1807.06757 (2018)

    Anderson, P., Chang, A., Chaplot, D.S., Dosovitskiy, A., Gupta, S., Koltun, V., Kosecka, J., Malik, J., Mottaghi, R., Savva, M., Zamir, A.R.: On evaluation of embodied navigation agents. arXiv preprint arXiv:1807.06757 (2018)

  2. [2]

    arXiv preprint arXiv:2001.10773 (2020)

    Cabon, Y., Murray, N., Humenberger, M.: Virtual KITTI 2. arXiv preprint arXiv:2001.10773 (2020)

  3. [3]

    In: CVPRW (2018)

    DeTone, D., Malisiewicz, T., Rabinovich, A.: SuperPoint: Self-supervised interest point detection and description. In: CVPRW (2018)

  4. [4]

    In: CVPR (2023)

    Edstedt, J., Athanasiadis, I., Wadenbäck, M., Felsberg, M.: DKM: Dense kernelized feature matching for geometry estimation. In: CVPR (2023)

  5. [5]

    In: CVPR (2024)

    Edstedt, J., Sun, Q., Bökman, G., Wadenbäck, M., Felsberg, M.: RoMa: Robust dense feature matching. In: CVPR (2024)

  6. [6]

    In: Proc

    Garg, S., Rana, K., Hosseinzadeh, M., Mares, L., Sünderhauf, N., Dayoub, F., Reid, I.: RoboHop: Segment-based topological map representation for open-world visual navigation. In: Proc. IEEE Int. Conf. Robotics and Automation (ICRA) (2024)

  7. [7]

    In: Proc

    Hughes, N., Chang, Y., Carlone, L.: Hydra: A real-time spatial perception system for 3D scene graph construction and optimization. In: Proc. Robotics: Science and Systems (RSS) (2022)

  8. [8]

    In: NeurIPS (2025)

    Jayanti, R., Agrawal, S., Garg, V., Tourani, S., Khan, M.H., Garg, S., Krishna, M.: SegMASt3R: Geometry grounded segment matching. In: NeurIPS (2025)

Show all 33 references
  1. [9]

    In: ICCV (2023) MuViSeg 15

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment anything. In: ICCV (2023) MuViSeg 15

  2. [10]

    In: ICCV (2023)

    Krantz, J., Gervet, T., Yadav, K., Wang, A., Paxton, C., Mottaghi, R., Batra, D., Malik, J., Lee, S., Chaplot, D.S.: Navigating to objects specified by images. In: ICCV (2023)

  3. [11]

    In: ECCV (2024)

    Leroy, V., Cabon, Y., Revaud, J.: Grounding image matching in 3D with MASt3R. In: ECCV (2024)

  4. [12]

    In: CVPR (2022)

    Li, R., Zhang, S., He, X.: SGTR: End-to-end scene graph generation with trans- former. In: CVPR (2022)

  5. [13]

    arXiv preprint arXiv:2511.10647 (2025)

    Lin, H., Chen, S., Liew, J.H., Chen, D.Y., Li, Z., Shi, G., Feng, J., Kang, B.: Depth anything 3: Recovering the visual space from any views. arXiv preprint arXiv:2511.10647 (2025)

  6. [14]

    In: ICCV (2023)

    Lindenberger, P., Sarlin, P.E., Pollefeys, M.: LightGlue: Local feature matching at light speed. In: ICCV (2023)

  7. [15]

    In: Proc

    Liu, Y., Lai, W., Zhao, Z., Xiong, Y., Zhu, J., Cheng, J., Xu, Y.: LiftFeat: 3D geometry-aware local feature matching. In: Proc. IEEE Int. Conf. Robotics and Automation (ICRA) (2025)

  8. [16]

    In: ICLR (2019)

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019)

  9. [17]

    TMLR (2024)

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W.,Howes,R.,Huang,P.Y.,Li,S.W.,Misra,I.,Rabbat,M.,Sharma,V.,Synnaeve, G., Xu, H., Jégou, H., Mairal, J., Labatut, ...

  10. [18]

    In: Proc

    Podgorski,S.,Garg,S.,Hosseinzadeh,M.,Mares,L.,Dayoub,F.,Reid,I.:TANGO: Traversability-aware navigation with local metric control for topological goals. In: Proc. IEEE Int. Conf. Robotics and Automation (ICRA). pp. 2399–2406 (2025)

  11. [19]

    In: Proc

    Ramakrishnan, S.K., Gokaslan, A., Wijmans, E., Maksymets, O., Clegg, A., Turner, J., Undersander, E., Galuba, W., Westbury, A., Chang, A.X., Savva, M., Zhao, Y., Batra, D.: Habitat-Matterport 3D dataset (HM3D): 1000 large-scale 3D environments for embodied AI. In: Proc. NeurIP...

  12. [20]

    In: ICCV (2021)

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

  13. [21]

    In: CVPR (2024)

    Ranzinger, M., Heinrich,G., Kautz, J.,Molchanov, P.:AM-RADIO: Agglomerative vision foundation model reduce all domains into one. In: CVPR (2024)

  14. [22]

    arXiv preprint arXiv:2408.00714 (2024)

    Ravi, N., Gabeur, V., Hu, Y.T., Hu, R., Ryali, C., Ma, T., Khedr, H., Rädle, R., Rolland, C., Gustafson, L., Mintun, E., Pan, J., Alwala, K.V., Carion, N., Wu, C.Y., Girshick, R., Dollár, P., Feichtenhofer, C.: SAM 2: Segment anything in images and videos. arXiv preprint arXiv...

  15. [23]

    In: CVPR (2020)

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

  16. [24]

    In: ICCV (2019)

    Savva, M., Kadian, A., Maksymets, O., Zhao, Y., Wijmans, E., Jain, B., Straub, J., Liu, J., Koltun, V., Malik, J., Parikh, D., Batra, D.: Habitat: A platform for embodied AI research. In: ICCV (2019)

  17. [25]

    In: ICLR (2024)

    Shen, X., Cai, Z., Yin, W., Müller, M., Li, Z., Wang, K., Chen, X., Wang, C.: GIM: Learning generalizable image matcher from internet videos. In: ICLR (2024)

  18. [26]

    Pacific Journal of Mathematics21(2), 343–348 (1967)

    Sinkhorn, R., Knopp, P.: Concerning nonnegative matrices and doubly stochastic matrices. Pacific Journal of Mathematics21(2), 343–348 (1967)

  19. [27]

    Fatykhoph et al

    Straub, J., Whelan, T., Ma, L., Chen, Y., Wijmans, E., Green, S., Engel, J.J., Mur- Artal, R., Ren, C., Verma, S., Clarkson, A., Yan, M., Budge, B., Yan, Y., Pan, X., Yon, J., Zou, Y., Leon, K., Carter, N., Briales, J., Gillingham, T., Mueggler, E., Pesqueira, L., Savva, M., B...

  20. [28]

    In: CVPR (2021)

    Sun, J., Shen, Z., Wang, Y., Bao, H., Zhou, X.: LoFTR: Detector-free local feature matching with transformers. In: CVPR (2021)

  21. [29]

    In: CVPR (2025)

    Wang, J., Chen, M., Karaev, N., Vedaldi, A., Rupprecht, C., Novotny, D.: VGGT: Visual geometry grounded transformer. In: CVPR (2025)

  22. [30]

    In: CVPR (2024)

    Wang, S., Leroy, V., Cabon, Y., Chidlovskii, B., Revaud, J.: DUSt3R: Geometric 3D vision made easy. In: CVPR (2024)

  23. [31]

    In: NeurIPS (2022)

    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)

  24. [32]

    In: ICCV (2023)

    Yeshwanth,C.,Liu,Y.C.,Nießner,M.,Dai,A.:ScanNet++:Ahigh-fidelitydataset of 3D indoor scenes. In: ICCV (2023)

  25. [33]

    arXiv preprint arXiv:2306.12156 (2023)

    Zhao, X., Ding, W., An, Y., Du, Y., Yu, T., Li, M., Tang, M., Wang, J.: Fast segment anything. arXiv preprint arXiv:2306.12156 (2023)

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.