Pith. sign in

REVIEW 5 major objections 2 minor 42 references

Symmetry-Constrained Multi-Scale Physics-Informed Neural Networks for Graphene Electronic Band Structure Prediction

T0 review · 5 major / 2 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read A neural network predicts graphene's band structure with a Dirac gap within 30.3 μeV of zero.

desk verdict The manuscript body is an unrelated cross-view localization paper, so the graphene PINN claims are checkable only as an abstract, making this record unrefereeable in its current form. read the letter →

arxiv 2508.10718 v1 pith:CJCWGUYQ submitted 2025-08-14 cond-mat.mtrl-sci cs.LGphysics.comp-ph

classification cond-mat.mtrl-scics.LGphysics.comp-ph
keywords graphenebandstructurephysics-informedneuralnetworkC6vsymmetryDiracconeBrillouinzonemulti-scalearchitecturetwo-dimensionalmaterials
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 a Symmetry-Constrained Multi-Scale Physics-Informed Neural Network (SCMS-PINN) learns graphene's electronic band structure directly from k-points while enforcing all twelve operations of the hexagonal C6v symmetry group. The proposed network splits the Brillouin zone into three ResNet-6 pathways — a K-head for the Dirac cone, an M-head for saddle points, and a General head for smooth interpolation — and progressively increases a Dirac constraint weight from 5.0 to 25.0 during training. Against 10,000 training k-points the authors report a 99.99% loss reduction, a validation loss of 0.0085, Dirac-point gaps within 30.3 μeV of zero, and average band errors of 53.9 meV (valence) and 40.5 meV (conduction). If these numbers hold, PINNs become a cheap, symmetry-exact surrogate for expensive band-structure calculations in 2D materials. The provided full text is a different manuscript, so the numerical claims are abstract-level assertions rather than verifiable results in this document.

What carries the argument

The load-bearing object is a three-headed ResNet-6 network fed by 31 physics-informed features per k-point, combined with systematic output averaging over the twelve operations of C6v, the hexagonal point group of graphene. Each head is specialized: K-head for the Dirac cone, M-head for the saddle point, and General head for smooth interpolation. The C6v averaging is what converts a generic neural regressor into a symmetry-respecting surrogate: every predicted energy is replaced by the mean over the symmetry orbit of the input k-point, so the final function is exactly invariant by construction. The trained function's role is to replace a full band-structure solver with a differentiable, fast evaluator.

What would settle it

Train the identical SCMS-PINN on 10,000 k-points generated by an independent density-functional or tight-binding code and check two numbers: validation loss near 0.0085 and a Dirac-point gap within 30.3 μeV of zero. Failing either, or finding that the symmetry-averaging step shifts energies by more than numerical precision, would refute the claimed exactness and accuracy.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes a recipe: encode a k-point as 31 physics-informed features, route it through three specialized ResNet-6 branches, and average the outputs over all twelve C6v operations so the predicted band energies are exactly invariant under the graphene point group. The K-head concentrates on the Dirac cone at K and K′, the M-head on the saddle points at M, and the General head fills in the rest of the Brillouin zone. Progressive Dirac constraint scheduling shifts the loss weight from 5.0 to 25.0 so that learning proceeds from global band topology to the local linear crossing. Training on 10,000 k-points over 300 epochs yields, in the paper's numbers, a training-loss drop from 34.597 to 0.003 and validation loss 0.0085, with the Dirac gap sitting 30.3 μeV from zero and average errors of 53.9 meV valence / 40.5 meV conduction across the zone.

Load-bearing premise

The reported accuracy is only as trustworthy as the reference band-structure data used to generate the 10,000 k-point training targets, and the supplied full text does not contain that calculation.

Editorial extensions

If this is right

  • If correct, a PINN can reproduce the Dirac cone with a near-zero gap without post-hoc correction, which would address the spurious-gap problem common in neural and plane-wave fits at band crossings.
  • The symmetry-averaging step gives an exact invariance guarantee, meaning the surrogate will never break C6v even if the underlying network is asymmetric.
  • The reported training cost of 10,000 k-points and 300 epochs suggests that cheap neural surrogates for 2D materials screening are within reach.
  • The multi-head specialization points to a general recipe: allocate network capacity near critical points rather than uniformly across the Brillouin zone.
  • The 40–54 meV average errors mean the surrogate is useful for band topology but not for meV-scale quantities other than the Dirac gap.

Reading between the lines

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

  • The 30.3 μeV Dirac gap and the roughly 50 meV average errors sit at very different scales, so a plausible reading is that the Dirac constraint forces extreme accuracy at K while leaving mid-zone errors about a thousand times larger; the headline number and the average error should be treated as separate claims.
  • Applied to lower-symmetry 2D materials such as h-BN or transition-metal dichalcogenides, the architecture would need a smaller symmetry group and a band-gap constraint in place of the Dirac constraint, a transfer the paper does not test.
  • A quantitative comparison between the symmetry-averaged network and an identical network without averaging would isolate exactly how much accuracy the C6v enforcement buys, which the paper does not report.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 2 minor

Summary. The abstract describes SCMS-PINN v35, a symmetry-constrained multi-scale physics-informed neural network for predicting graphene electronic band structures. It claims that three specialized ResNet-6 pathways act on 31 physics-informed features, that training on 10,000 k-points over 300 epochs reduces the training loss from 34.597 to 0.003, that the model predicts Dirac point gaps within 30.3 μeV of theoretical zero, and that average errors are 53.9 meV (valence) and 40.5 meV (conduction) across the Brillouin zone. The supplied full text, however, is the paper 'ViewBridge: Revisiting Cross-View Localization from Image Matching' (arXiv:2508.10716v2). It contains no graphene, band structure, PINN, Dirac, C6v, or symmetry-averaging content. None of the methods, data, architecture details, training/validation splits, reference calculations, or error definitions needed to support the abstract's claims appear anywhere in the manuscript.

Significance. If substantiated, the abstract's claims would be of interest to the condensed-matter and machine-learning communities: a physics-informed network that enforces C6v symmetry and achieves roughly 50 meV band-structure errors while resolving the Dirac gap to tens of microelectronvolts would be a noteworthy methodological contribution. However, the manuscript as submitted supplies none of the supporting evidence. There is no reference band-structure calculation, no architecture diagram or hyperparameter table, no dataset description, no code or reproducibility artifact, and no derivation of the symmetry-averaging procedure. The reported numbers are bare assertions. There is also a conceptual risk: if symmetry is enforced by averaging the network output over the twelve C6v operations, the Dirac-point degeneracy is imposed by construction, making the 30.3 μeV 'gap to theoretical zero' a property of the constraint rather than an independent prediction. Because the full text is an unrelated paper, the central claim is not checkable in any form.

major comments (5)
  1. [Full text (all sections)] The submitted manuscript body is the paper 'ViewBridge: Revisiting Cross-View Localization from Image Matching' (arXiv:2508.10716v2). It contains no derivation, architecture description, data description, training details, or evaluation relating to SCMS-PINN, graphene, band structures, Dirac points, or C6v symmetry. Every load-bearing element of the abstract's central claim—architecture, reference band-structure calculation, training/validation split, loss definitions, and error metrics—is absent from the manuscript. The central claim is therefore unsupported as submitted.
  2. [Abstract] The abstract reports precise numerical results (training loss reduction from 34.597 to 0.003, validation loss 0.0085, Dirac gap within 30.3 μeV, valence error 53.9 meV, conduction error 40.5 meV) but gives no definition of these errors. It is not specified whether the errors are mean absolute errors, over which k-point sets they are computed, whether the validation points are distinct from the 10,000 training points, or what reference calculation defines 'theoretical zero.' Without these definitions, the numbers cannot be reproduced or interpreted.
  3. [Abstract] The claim that 'All twelve C6v operations are enforced through systematic averaging, guaranteeing exact symmetry preservation' raises a circularity concern. If the network output is symmetrized by group-averaging, the Dirac-point degeneracy at the K/K' points is built into the model by construction. The reported 30.3 μeV deviation from 'theoretical zero' is then a measure of the averaging/fitting procedure, not an independently predicted physical quantity. The abstract provides no proof or sensitivity analysis that the averaging preserves the learned band topology without distorting it.
  4. [Abstract] The description of the 'progressive Dirac constraint scheduling' that increases a weight parameter from 5.0 to 25.0 is not accompanied by any details of the loss function, the constraint term, or the scheduling rule. Because the full text is unrelated to the abstract, there is no way to determine whether the reported loss trajectory and error metrics come from the claimed model, a different model, or any model at all.
  5. [Abstract and title] The title and abstract describe a materials-science manuscript, while the full text is a computer-vision paper. The manuscript therefore fails the basic requirement that the stated claims be supported by the presented methods and results. This is not a local or cosmetic issue; the entirety of the supporting evidence is missing.
minor comments (2)
  1. [Abstract] The abstract is a standalone claim with no references to prior work on symmetry-constrained neural networks, graphene band-structure models, or the reference electronic-structure method, so the reader cannot locate the code, data, or context needed to assess the contribution.
  2. [Abstract] The label 'v35' for the model is unexplained; without a methods section, the significance of this version identifier is unclear.

Circularity Check

1 steps flagged · score 6.0 of 10

The Dirac-gap 'prediction' is built in by C6v averaging and a scheduled Dirac constraint; the full text is an unrelated cross-view localization paper.

  1. fitted input called prediction [Abstract (paras 3-4): Dirac constraint scheduling and C6v averaging sentences.]
    "Progressive Dirac constraint scheduling systematically increases the weight parameter from 5.0 to 25.0... The model predicts Dirac point gaps within 30.3 $\mu$eV of theoretical zero... All twelve C$_{6v}$ operations are enforced through systematic averaging, guaranteeing exact symmetry preservation."

    The 30.3 μeV 'prediction' is manufactured by the model's own constraints. Averaging outputs over the twelve C6v operations projects every prediction onto the fully symmetric subspace; at the Dirac point K, C6v symmetry plus time-reversal forces the two π bands to be exactly degenerate, so the symmetrized output gap at K is identically zero up to floating-point noise. Additionally, the scheduled loss term with weight rising to 25.0 explicitly penalizes any nonzero gap at K during training. The reported residual is therefore the tolerance of the symmetry projection and constraint, not an independent physical prediction. The abstract offers no unsymmetrized or unconstrained model for comparison, and the full text contains no methods or equations that would allow the claim to be checked.

full rationale

The only checkable derivation chain is the abstract, since the supplied full text is arXiv:2508.10716v2, titled 'ViewBridge: Revisiting Cross-View Localization from Image Matching'—a different paper with no graphene, PINN, band-structure, or C6v content. Within the abstract, the headline success at the Dirac point is circular: the output is averaged over all twelve C6v operations, which by construction makes the K-point π bands degenerate, and the loss schedule explicitly raises a Dirac-constraint weight to 25.0, penalizing any nonzero gap. Thus the 30.3 μeV gap is the residual of an imposed symmetry and training objective rather than an independent prediction. The BZ-averaged errors (53.9 and 40.5 meV) are not forced by the symmetry and would be meaningful if the reference band-structure calculation and train/validation split were described, but no such methods exist in the supplied text. Because the central 'Dirac gap prediction' reduces by construction to the model's constraints, the circularity score is 6; the full-text mismatch and missing methods are separate integrity concerns, not additional circularity.

Assumptions & free parameters 2 free parameters · 2 assumptions · 0 invented entities

Extracted from the abstract only; the full text is a different paper. Free parameters and assumptions are limited to what the abstract states explicitly.

free parameters (2)
  • Dirac constraint weight = 5.0 to 25.0 (schedule)
    Abstract: 'Progressive Dirac constraint scheduling systematically increases the weight parameter from 5.0 to 25.0.' This is a hand-tuned schedule that controls the strength of the Dirac-point constraint.
  • Number of physics-informed features = 31
    Abstract: 'operating on 31 physics-informed features extracted from k-points.' The specific features are not enumerated; the count is an architectural choice.
assumptions (2)
  • domain assumption Averaging network outputs over the C6v group exactly preserves the band structure symmetry.
    Abstract: 'All twelve C6v operations are enforced through systematic averaging, guaranteeing exact symmetry preservation.' This assumes group averaging is both necessary and sufficient for physical symmetry.
  • domain assumption The 10,000 k-point training labels are accurate reference band energies.
    Abstract: 'Training on 10,000 k-points over 300 epochs.' No source of band-structure labels (e.g., DFT or tight-binding) is given in the abstract.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Symmetry-Constrained Multi-Scale Physics-Informed Neural Networks for Graphene Electronic Band Structure Prediction." pith.science (2026). https://pith.science/paper/CJCWGUYQ

@misc{pith2026250810718,
  author       = {Pith},
  title        = {Pith review of: Symmetry-Constrained Multi-Scale Physics-Informed Neural Networks for Graphene Electronic Band Structure Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CJCWGUYQ}},
  note         = {Machine review of arXiv:2508.10718}
}
abstract

Accurate prediction of electronic band structures in two-dimensional materials remains a fundamental challenge, with existing methods struggling to balance computational efficiency and physical accuracy. We present the Symmetry-Constrained Multi-Scale Physics-Informed Neural Network (SCMS-PINN) v35, which directly learns graphene band structures while rigorously enforcing crystallographic symmetries through a multi-head architecture. Our approach introduces three specialized ResNet-6 pathways -- K-head for Dirac physics, M-head for saddle points, and General head for smooth interpolation -- operating on 31 physics-informed features extracted from k-points. Progressive Dirac constraint scheduling systematically increases the weight parameter from 5.0 to 25.0, enabling hierarchical learning from global topology to local critical physics. Training on 10,000 k-points over 300 epochs achieves 99.99\% reduction in training loss (34.597 to 0.003) with validation loss of 0.0085. The model predicts Dirac point gaps within 30.3 $\mu$eV of theoretical zero and achieves average errors of 53.9 meV (valence) and 40.5 meV (conduction) across the Brillouin zone. All twelve C$_{6v}$ operations are enforced through systematic averaging, guaranteeing exact symmetry preservation. This framework establishes a foundation for extending physics-informed learning to broader two-dimensional materials for accelerated discovery.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 34 canonical work pages

  1. [1]

    Map-free visual relocalization: Metric pose relative to a single image

    Eduardo Arnold, Jamie Wynn, Sara Vicente, Guillermo Garcia-Hernando, Aron Monszpart, Victor Prisacariu, Dani- yar Turmukhambetov, and Eric Brachmann. Map-free visual relocalization: Metric pose relative to a single image. In European Conference on Computer Vision, pages 690–708. Springer, 2022. 1

  2. [2]

    Matching 2d images in 3d: Metric relative pose from metric correspondences

    Axel Barroso-Laguna, Sowmya Munukutla, Victor Adrian Prisacariu, and Eric Brachmann. Matching 2d images in 3d: Metric relative pose from metric correspondences. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4852–4863, 2024. 2

  3. [3]

    Sam- ple4geo: Hard negative sampling for cross-view geo- localisation

    Fabian Deuser, Konrad Habel, and Norbert Oswald. Sam- ple4geo: Hard negative sampling for cross-view geo- localisation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 16847–16856, 2023. 2

  4. [4]

    Roma: Robust dense feature matching

    Johan Edstedt, Qiyu Sun, Georg B¨okman, M˚arten Wadenb¨ack, and Michael Felsberg. Roma: Robust dense feature matching. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19790–19800, 2024. 2, 7, 8

  5. [5]

    3d wireframe model reconstruction of buildings from multi-view images using neural implicit fields.Automation in Construction, 174:106145, 2025

    Weiwei Fan, Xinyi Liu, Yongjun Zhang, Dong Wei, Haoyu Guo, and Dongdong Yue. 3d wireframe model reconstruction of buildings from multi-view images using neural implicit fields.Automation in Construction, 174:106145, 2025. 3

  6. [6]

    Uncertainty-aware vision-based metric cross-view geolocalization

    Florian Fervers, Sebastian Bullinger, Christoph Bodensteiner, Michael Arens, and Rainer Stiefelhagen. Uncertainty-aware vision-based metric cross-view geolocalization. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21621–21631, 2023. 2

  7. [7]

    Slicematch: Geometry-guided aggregation for cross-view pose estimation

    Ted Lentsch, Zimin Xia, Holger Caesar, and Julian FP Kooij. Slicematch: Geometry-guided aggregation for cross-view pose estimation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17225– 17234, 2023. 1, 2, 6

  8. [8]

    Ground- ing image matching in 3d with mast3r

    Vincent Leroy, Yohann Cabon, and J´erˆome Revaud. Ground- ing image matching in 3d with mast3r. InEuropean Confer- ence on Computer Vision, pages 71–91. Springer, 2024. 2, 7

Show all 42 references
  1. [9]

    Cross-view geolocalization and disaster mapping with street- view and vhr satellite imagery: A case study of hurricane ian

    Hao Li, Fabian Deuser, Wenping Yin, Xuanshu Luo, Paul Walther, Gengchen Mai, Wei Huang, and Martin Werner. Cross-view geolocalization and disaster mapping with street- view and vhr satellite imagery: A case study of hurricane ian. ISPRS Journal of Photogrammetry and Remote Sen...

  2. [10]

    Bevformer: learning bird’s-eye-view representation from lidar-camera via spatiotemporal transformers.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chong- hao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: learning bird’s-eye-view representation from lidar-camera via spatiotemporal transformers.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3, 6

  3. [11]

    Stereoinr: Cross-view geometry consistent stereo super resolution with implicit neural representation

    Yi Liu, Xinyi Liu, Panwang Xia, Qiong Wu, Yi Wan, and Yongjun Zhang. Stereoinr: Cross-view geometry consistent stereo super resolution with implicit neural representation. arXiv preprint arXiv:2505.05509, 2025. 3

  4. [12]

    Gleam: Learning to match and explain in cross-view geo-localization.arXiv preprint arXiv:2509.07450, 2025

    Xudong Lu, Zhi Zheng, Yi Wan, Yongxiang Yao, Annan Wang, Renrui Zhang, Panwang Xia, Qiong Wu, Qingyun Li, Weifeng Lin, et al. Gleam: Learning to match and explain in cross-view geo-localization.arXiv preprint arXiv:2509.07450, 2025. 2

  5. [13]

    1 year, 1000 km: The oxford robotcar dataset.The International Journal of Robotics Research, 36(1):3–15, 2017

    Will Maddern, Geoffrey Pascoe, Chris Linegar, and Paul Newman. 1 year, 1000 km: The oxford robotcar dataset.The International Journal of Robotics Research, 36(1):3–15, 2017. 1

  6. [14]

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

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis.Communications of the ACM, 65(1):99–106, 2021. 3

  7. [15]

    Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2023

    Maxime Oquab, Timoth´ee Darcet, Th´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2023. 3

  8. [16]

    Superglue: Learning feature match- ing with graph neural networks

    Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature match- ing with graph neural networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recog- nition, pages 4938–4947, 2020. 2, 7, 8

  9. [17]

    Beyond cross-view image re- trieval: Highly accurate vehicle localization using satellite image

    Yujiao Shi and Hongdong Li. Beyond cross-view image re- trieval: Highly accurate vehicle localization using satellite image. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 17010–17020,

  10. [18]

    Spatial- aware feature aggregation for image based cross-view geo- localization.Advances in Neural Information Processing Systems, 32, 2019

    Yujiao Shi, Liu Liu, Xin Yu, and Hongdong Li. Spatial- aware feature aggregation for image based cross-view geo- localization.Advances in Neural Information Processing Systems, 32, 2019. 2

  11. [19]

    Accurate 3-dof camera geo-localization via ground-to-satellite image matching.IEEE transactions on pattern analysis and machine intelligence, 45(3):2682–2697,

    Yujiao Shi, Xin Yu, Liu Liu, Dylan Campbell, Piotr Koniusz, and Hongdong Li. Accurate 3-dof camera geo-localization via ground-to-satellite image matching.IEEE transactions on pattern analysis and machine intelligence, 45(3):2682–2697,

  12. [20]

    Boosting 3-dof ground-to-satellite camera localiza- tion accuracy via geometry-guided cross-view transformer

    Yujiao Shi, Fei Wu, Akhil Perincherry, Ankit V ora, and Hong- dong Li. Boosting 3-dof ground-to-satellite camera localiza- tion accuracy via geometry-guided cross-view transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21516–21526, 2...

  13. [21]

    Weakly-supervised camera localization by ground-to-satellite image registration

    Yujiao Shi, Hongdong Li, Akhil Perincherry, and Ankit V ora. Weakly-supervised camera localization by ground-to-satellite image registration. InEuropean Conference on Computer Vision, pages 39–57. Springer, 2024

  14. [22]

    Learning dense flow field for highly-accurate cross-view camera localiza- tion.Advances in Neural Information Processing Systems, 36:70612–70625, 2023

    Zhenbo Song, Jianfeng Lu, Yujiao Shi, et al. Learning dense flow field for highly-accurate cross-view camera localiza- tion.Advances in Neural Information Processing Systems, 36:70612–70625, 2023. 1, 2, 6 9

  15. [23]

    Loftr: Detector-free local feature matching with transformers

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

  16. [24]

    Aerialmegadepth: Learn- ing aerial-ground reconstruction and view synthesis

    Khiem Vuong, Anurag Ghosh, Deva Ramanan, Srinivasa Narasimhan, and Shubham Tulsiani. Aerialmegadepth: Learn- ing aerial-ground reconstruction and view synthesis. InPro- ceedings of the Computer Vision and Pattern Recognition Conference, pages 21674–21684, 2025. 2, 7, 8

  17. [25]

    Vggt: Visual geometry grounded transformer

    Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 5294–5306, 2025. 2

  18. [26]

    Dust3r: Geometric 3d vision made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 20697–20709, 2024. 2

  19. [27]

    Fine-grained cross-view geo-localization using a correlation-aware homography estimator.Advances in Neural Information Processing Systems, 36:5301–5319, 2023

    Xiaolong Wang, Runsen Xu, Zhuofan Cui, Zeyu Wan, and Yu Zhang. Fine-grained cross-view geo-localization using a correlation-aware homography estimator.Advances in Neural Information Processing Systems, 36:5301–5319, 2023. 1, 2, 6

  20. [28]

    Camp: A cross-view geo-localization method using contrastive attributes mining and position-aware partitioning.IEEE Transactions on Geo- science and Remote Sensing, 2024

    Qiong Wu, Yi Wan, Zhi Zheng, Yongjun Zhang, Guang- shuai Wang, and Zhenyang Zhao. Camp: A cross-view geo-localization method using contrastive attributes mining and position-aware partitioning.IEEE Transactions on Geo- science and Remote Sensing, 2024. 2

  21. [29]

    Cross-view image set geo-localization.arXiv preprint arXiv:2412.18852, 2024

    Qiong Wu, Panwang Xia, Lei Yu, Yi Liu, Mingtao Xiong, Liheng Zhong, Jingdong Chen, Ming Yang, Yongjun Zhang, and Yi Wan. Cross-view image set geo-localization.arXiv preprint arXiv:2412.18852, 2024

  22. [30]

    Enhancing cross-view geo-localization with domain alignment and scene consistency.IEEE Transactions on Cir- cuits and Systems for Video Technology, 2024

    Panwang Xia, Yi Wan, Zhi Zheng, Yongjun Zhang, and Jiwei Deng. Enhancing cross-view geo-localization with domain alignment and scene consistency.IEEE Transactions on Cir- cuits and Systems for Video Technology, 2024

  23. [31]

    Cross-view geo-localization with panoramic street- view and vhr satellite imagery in decentrality settings.ISPRS Journal of Photogrammetry and Remote Sensing, 227:1–11,

    Panwang Xia, Lei Yu, Yi Wan, Qiong Wu, Peiqi Chen, Liheng Zhong, Yongxiang Yao, Dong Wei, Xinyi Liu, Lixiang Ru, et al. Cross-view geo-localization with panoramic street- view and vhr satellite imagery in decentrality settings.ISPRS Journal of Photogrammetry and Remote Sensing...

  24. [32]

    Fgˆ 2: Fine-grained cross- view localization by fine-grained feature matching

    Zimin Xia and Alexandre Alahi. Fgˆ 2: Fine-grained cross- view localization by fine-grained feature matching. InPro- ceedings of the Computer Vision and Pattern Recognition Conference, pages 6362–6372, 2025. 1, 2, 3, 5, 6, 7, 8

  25. [33]

    Visual cross-view metric localization with dense uncertainty estimates

    Zimin Xia, Olaf Booij, Marco Manfredi, and Julian FP Kooij. Visual cross-view metric localization with dense uncertainty estimates. InEuropean Conference on Computer Vision, pages 90–106. Springer, 2022. 1, 2

  26. [34]

    Convolutional cross-view pose estimation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(5):3813–3831, 2023

    Zimin Xia, Olaf Booij, and Julian FP Kooij. Convolutional cross-view pose estimation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(5):3813–3831, 2023. 1, 2, 6

  27. [35]

    Adapting fine-grained cross-view localization to areas without fine ground truth

    Zimin Xia, Yujiao Shi, Hongdong Li, and Julian FP Kooij. Adapting fine-grained cross-view localization to areas without fine ground truth. InEuropean Conference on Computer Vision, pages 397–415. Springer, 2024. 2

  28. [36]

    Fine-grained cross-view localization via local feature matching and monoc- ular depth priors.arXiv preprint arXiv:2509.09792, 2025

    Zimin Xia, Chenghao Xu, and Alexandre Alahi. Fine-grained cross-view localization via local feature matching and monoc- ular depth priors.arXiv preprint arXiv:2509.09792, 2025. 2, 4

  29. [37]

    Depth anything v2.Advances in Neural Information Processing Systems, 37: 21875–21911, 2024

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiao- gang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2.Advances in Neural Information Processing Systems, 37: 21875–21911, 2024. 4

  30. [38]

    Sg-bev: satellite-guided bev fusion for cross-view semantic segmentation

    Junyan Ye, Qiyan Luo, Jinhua Yu, Huaping Zhong, Zhimeng Zheng, Conghui He, and Weijia Li. Sg-bev: satellite-guided bev fusion for cross-view semantic segmentation. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27748–27757, 2024. 1

  31. [39]

    Nerfortho: Ortho- graphic projection images generation based on neural radi- ance fields.International Journal of Applied Earth Observa- tion and Geoinformation, 136:104378, 2025

    Dongdong Yue, Xinyi Liu, Yi Wan, Yongjun Zhang, Maoteng Zheng, Weiwei Fan, and Jiachen Zhong. Nerfortho: Ortho- graphic projection images generation based on neural radi- ance fields.International Journal of Applied Earth Observa- tion and Geoinformation, 136:104378, 2025. 3

  32. [40]

    Cross-view localization via redundant sliced observations and a-contrario validation.arXiv preprint arXiv:2508.05369,

    Yongjun Zhang, Mingtao Xiong, Yi Wan, and Gui-Song Xia. Cross-view localization via redundant sliced observations and a-contrario validation.arXiv preprint arXiv:2508.05369,

  33. [41]

    Vigor: Cross- view image geo-localization beyond one-to-one retrieval

    Sijie Zhu, Taojiannan Yang, and Chen Chen. Vigor: Cross- view image geo-localization beyond one-to-one retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 3640–3649, 2021. 5, 6

  34. [42]

    Deformable detr: Deformable trans- formers for end-to-end object detection.arXiv preprint arXiv:2010.04159, 2020

    Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable trans- formers for end-to-end object detection.arXiv preprint arXiv:2010.04159, 2020. 3 10 perGlue [16], RoMa [4], Aerial-Megadepth [24], FG2 [32], and our approach. Green and r...

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.