Pith. sign in

REVIEW 3 major objections 5 minor 46 references

VMatcher: State-Space Semi-Dense Local Feature Matching

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read VMatcher claims a hybrid Mamba-Transformer design delivers ELoFTR-level matching accuracy at lower runtime.

desk verdict VMatcher is a credible engineering combination that roughly matches ELoFTR accuracy; the speed edge is real but limited to high resolutions, and the accuracy claim is within noise. read the letter →

arxiv 2507.23371 v1 pith:ZZ3NKBKQ submitted 2025-07-31 cs.CV

Ali Youssef This is my paper · ORCID
classification cs.CV
keywords featurematchingsemi-denseMambastate-spacemodelstransformerattentionrelativeposeestimationvisuallocalizationefficientinference
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

VMatcher is a semi-dense image matcher that replaces most Transformer attention layers with Mamba-style selective state-space layers while keeping a compact downsampled attention module. The paper argues this hybrid reaches the accuracy of detector-free matchers such as ELoFTR and LoFTR at a lower inference cost: on MegaDepth relative pose estimation, VMatcher-B reports AUC@10° 72.2 in 87.56 ms versus ELoFTR's 72.1 in 100.25 ms, and the optimized tiny variant runs at 55.94 ms. The practical stake is that dense, detector-free matching, which normally carries a heavy compute penalty, can approach the speed of sparse matchers without losing accuracy. The ablations support the hybrid claim by showing that neither the Mamba branch nor the transformer branch alone matches the full model.

What carries the argument

The load-bearing object is the hybrid Mamba-Transformer stage: MambaVision blocks (a selective state-space model with a symmetric convolution branch) give linear-time sequence processing, interleaved with gMLP layers and a Downsampled-Transformer module whose self- and cross-attention run on feature maps reduced by bilinear interpolation before being upsampled. The downsampling makes attention cheap enough to keep, while the state-space layers carry long-range context; the paper's ablation study shows removing either branch substantially lowers MegaDepth AUC. A VGG-style backbone supplies coarse and fine feature maps, and an ELoFTR-style two-stage patch refinement converts coarse matches into sub-pixel correspondences.

What would settle it

Rerun the MegaDepth pose evaluation with a single RANSAC inlier threshold shared by all methods, or with thresholds chosen on a validation split rather than the test split, and check whether VMatcher-B keeps its AUC@10° of 72.2 against ELoFTR's 72.1; if the ordering reverses beyond the reported roughly 1.2% spread, the parity claim fails. As a second check, measure end-to-end runtime at 480×640 with the 0.8 ms-per-layer Mamba overhead included; if the gap against ELoFTR nearly vanishes there, the efficiency claim is resolution-dependent.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that a hybrid Mamba-Transformer block is a viable global-context engine for semi-dense matching. MambaVision layers process long flattened feature sequences in linear time, a Downsampled-Transformer computes self- and cross-attention on bilinearly downsampled maps to keep the attention cost low, and gMLP layers provide local mixing; together with a lightweight VGG-style extractor and ELoFTR-style two-stage fine refinement, this reproduces or slightly exceeds the pose-estimation accuracy of ELoFTR while running faster. The paper also reports that a lightweight backbone outperforms a RepVGG backbone in their pipeline, that rotary position embeddings can be dropped from self-attention without accuracy loss, and that bidirectional Mamba scans add runtime without meaningful gains.

Load-bearing premise

The parity claim assumes that tuning the robust-fitting inlier threshold separately for every model on each test set favors all methods equally; if that tuning helped VMatcher, its accuracy showing could shrink.

Editorial extensions

If this is right

  • Detector-free matching can run at sparse-like speeds at high resolutions, since Mamba's per-layer overhead is roughly constant while attention cost grows with sequence length.
  • The reported MegaDepth numbers imply that a matching pipeline can keep roughly 72% AUC@10° on outdoor pose estimation while cutting inference time by 12–45% relative to ELoFTR.
  • Removing the post-attention MLP and rotary position embeddings are reported as safe simplifications, pointing to leaner attention blocks in future matchers.
  • The optimized variants' near-sparse runtimes make semi-dense correspondences more plausible for real-time localization and mapping loops.

Reading between the lines

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

  • The accuracy comparison rests on per-model RANSAC threshold tuning on each test set; if that tuning is not equally neutral to all methods, the claimed parity could be weaker than the tables suggest.
  • The reported speed advantage is mostly a high-resolution effect: at 480×640 the Mamba overhead of about 0.8 ms per layer leaves VMatcher close to ELoFTR, so low-resolution users may not see the headline gain.
  • Because training used only MegaDepth, a direct test would fine-tune on indoor or cross-domain data; the ScanNet results already hint that indoor gains are smaller than outdoor ones.
  • A pure-Mamba matcher with attention only at coarse cross-attention might be enough, since the paper's ablation without RoPE suggests positional encoding can be pruned further.
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

3 major / 5 minor

Summary. The paper proposes VMatcher, a hybrid Mamba-Transformer architecture for semi-dense local feature matching. The method uses a lightweight VGG-style extractor, MambaVision layers for efficient sequence modeling, a downsampled transformer for self-/cross-attention, and a coarse-to-fine refinement module adapted from ELoFTR. The authors present two configurations (Base and Tiny) plus optimized variants, and evaluate on HPatches, MegaDepth, ScanNet, Aachen v1.1, and InLoc, reporting accuracy and runtime comparisons against sparse and semi-dense baselines. The central claim is that VMatcher variants achieve accuracy on par with or better than semi-dense methods such as ELoFTR while providing runtime advantages, especially at higher resolutions.

Significance. If the empirical claims hold, VMatcher is a useful practical contribution: it combines a linear-complexity state-space model with a downsampled attention mechanism to reduce inference time in semi-dense matching, and the paper includes ablations that support the main architectural choices. The code is released, the training and evaluation details are mostly transparent, and the architecture equations are standard. The practical impact, however, hinges on the robustness of the reported accuracy and runtime comparisons, which are currently undermined by the evaluation protocol and by the resolution-dependent nature of the speed advantage.

major comments (3)
  1. [§5, Tables 1–3] The evaluation protocol tunes RANSAC's inlier threshold for each model on each test dataset and reports no error bars or threshold-sensitivity analysis. Because the reported accuracy differences between VMatcher and ELoFTR are typically 0.1–0.6 AUC points, they are of the same order as the granularity of this per-model, test-set-dependent tuning, so the claim that VMatcher 'delivers improved performance' (Sec. 5.2) is not distinguished from evaluation noise. The authors should either fix the RANSAC threshold across models, select it on a validation split, or provide error bars and a threshold sweep.
  2. [§5.2, Table 3, Fig. 4, Appendix H] The runtime advantage is not universal: on ScanNet at 480x640, VMatcher-B takes 24.52 ms versus 23.92 ms for ELoFTR (Table 3), and Appendix H concedes that for small sequences Mamba's runtime is comparable to Transformer-based models. The abstract and Sec. 5.2 state that VMatcher 'delivers improved performance and runtime speed advantages' over semi-dense methods without this qualification. The paper should state the resolution/sequence-length regime in which the efficiency claim holds and characterize the crossover point.
  3. [§5.2, Table 2] The paper states that VMatcher variants 'outperform all sparse and semi-dense methods' on MegaDepth, but the observed differences in AUC@10° are around 0.3 points (e.g., 72.2 for VMatcher-B vs. 72.1 for ELoFTR) and are well within the noise level expected from single runs without error bars. The claim of 'outperform' should be softened to 'on par' unless statistical significance is demonstrated or multiple seeds are reported.
minor comments (5)
  1. [Eq. (5)] Equation (5) is garbled: 'yi = LX i=0 C ¯AL−i ¯Bxi' should be a summation over an index with a clear upper limit, and the convolution kernel definition that follows is ambiguous.
  2. [§5.2, text near Table 3] The sentence 'VMatcher-B runtime is similar to ELoFTR [35]' cites reference [35] (LoFTR), but ELoFTR is reference [41]; the citation should be corrected.
  3. [§5.1, Table 1] The text says 'VMatcher-B matches ELoFTR in runtime', but Table 1 reports 35.43 ms for VMatcher-B versus 34.48 ms for ELoFTR; 'matches' is imprecise since VMatcher-B is slightly slower.
  4. [Figure 4 and Appendix A] The runtime breakdown legends include a component named 'ConcatMLP', but this term is never defined in the main text or the appendix; the layer patterns in Appendix A use 'G' for gMLP, so the legend should be aligned with the naming used elsewhere.
  5. [Appendix D.1 and Sec. 5.4] References [6] and [12] are both cited as 'Mamba' in Sec. 5.4, but [6] is 'Transformers are SSMs' rather than the Mamba paper; only [12] should be attributed to the Mamba architecture.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: VMatcher's claims are benchmark-driven and its architecture choices are empirically ablated rather than derived from the target results.

full rationale

VMatcher's central claims—a hybrid Mamba-Transformer architecture, accuracy on par with ELoFTR/LoFTR, and runtime gains—are supported by external benchmarks (HPatches, MegaDepth, ScanNet, Aachen, InLoc) and measured runtimes. No load-bearing step reduces to its inputs by construction. The MambaVision module is adopted from prior external work [15]; the fine-level refinement is explicitly 'Adapted from ELoFTR [41]', an external method; and the DS-Transformer design choices (downsampling, omitting the MLP, removing RoPE) are justified by ablations in Tab. 6 and design experiments in Appendix D, not by the target results. The only questionable evaluation practice is per-model RANSAC threshold tuning on test sets (Sec. 5) without error bars, but this is a benchmark-protocol caveat, not a circular derivation: the reported AUC numbers are not equal to the tuned threshold by construction, and all baselines receive the same treatment. The paper's own Appendix H concedes that Mamba runtime is comparable to Transformer-based models for shorter sequences and favorable for longer sequences, which is an honest limitation of the efficiency claim rather than circular reasoning. There are no self-citations, and no uniqueness theorem is imported. Therefore no circular step is identified.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The architecture relies on established building blocks (MambaVision, ELoFTR fine refinement) and standard evaluation datasets. The only hand-set hyperparameters are loss weights, downsampling factor, and RANSAC thresholds. No new entities are postulated.

free parameters (4)
  • Loss weights alpha and beta = 1.0 and 0.25
    Set by hand in Eq. 12; affect the optimization but are not fitted to the target benchmarks.
  • Attention downsampling factor = 4
    Bilinear interpolation factor before attention in Sec. 3.3; chosen by design with ablations in Appendix D.1.
  • RANSAC inlier threshold = tuned per model per dataset
    Tuned on each test dataset in Sec. 5; directly affects reported accuracy numbers.
  • Coarse matching threshold tau = not reported
    Threshold for mutual nearest neighbour matching in Sec. 3.4; value not specified in the paper.
assumptions (5)
  • domain assumption MambaVision is a valid backbone for 2D image feature extraction.
    Used as the core Mamba layer in Sec. 3.2.2, relying on prior work [15] rather than on a derivation in this paper.
  • domain assumption ELoFTR's two-stage fine refinement transfers to VMatcher's features.
    Sec. 3.5 states the module is adapted from ELoFTR; the paper does not re-derive or independently validate it.
  • standard math Standard SSM discretization and convolution formulas are correct.
    Sec. 3.2.1 equations (1) to (6) use standard SSM theory; no novel math is introduced.
  • domain assumption Training only on MegaDepth generalizes to other test datasets.
    Appendix B trains on MegaDepth only; evaluation on HPatches, ScanNet, Aachen, and InLoc assumes transfer.
  • domain assumption RoPE can be omitted because Mamba encodes positional information.
    Sec. 3.3 hypothesizes this; ablation row g in Table 6 supports it empirically.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VMatcher: State-Space Semi-Dense Local Feature Matching." pith.science (2026). https://pith.science/paper/ZZ3NKBKQ

@misc{pith2026250723371,
  author       = {Pith},
  title        = {Pith review of: VMatcher: State-Space Semi-Dense Local Feature Matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZZ3NKBKQ}},
  note         = {Machine review of arXiv:2507.23371}
}
read the original abstract

This paper introduces VMatcher, a hybrid Mamba-Transformer network for semi-dense feature matching between image pairs. Learning-based feature matching methods, whether detector-based or detector-free, achieve state-of-the-art performance but depend heavily on the Transformer's attention mechanism, which, while effective, incurs high computational costs due to its quadratic complexity. In contrast, Mamba introduces a Selective State-Space Model (SSM) that achieves comparable or superior performance with linear complexity, offering significant efficiency gains. VMatcher leverages a hybrid approach, integrating Mamba's highly efficient long-sequence processing with the Transformer's attention mechanism. Multiple VMatcher configurations are proposed, including hierarchical architectures, demonstrating their effectiveness in setting new benchmarks efficiently while ensuring robustness and practicality for real-time applications where rapid inference is crucial. Source Code is available at: https://github.com/ayoussf/VMatcher

Figures

Figures reproduced from arXiv: 2507.23371 by the authors.

Figure 1
Figure 1. VMatcher pipeline overview. 1) VGG-style backbone extracts multi-scale feature maps from images IA and IB, including coarse feature maps F˜CA and F˜CB at 1 8 resolution. 2) Hybrid Mamba-Transformer module processes these for enhanced discrimination. 3) Coarse-level matches {Mc} obtained via correlation and MNN matching. 4) Fine-level feature maps are generated by fusing F˜CA and F˜CB with 1 4 and 1 2 resolution back… view at source ↗
Figure 2
Figure 2. Mamba Block Comparison: (a) Vanilla Mamba [ [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Downsampled Transformer illustration. ability map, as in [11, 35, 38, 41]. Coarse-level matches {Mc} are obtained using mutual nearest neighbour (MNN) and a threshold τ . 3.5. Fine-Level Extraction and Refinement Adapted from ELoFTR [41], the fine-level refinement mod￾ule efficiently achieves sub-pixel precision through two components: Fine Feature Extraction. The transformed coarse feature vectors F˜t CA and F˜t CB… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Runtime Breakdown. A runtime breakdown of VMatcher’s pipeline across multiple image resolutions shows efficiency gains over ELoFTR [41] as the resolution increases. Matcher Type Method ScanNet Dataset Time (ms) AUC@5 ◦ / AUC@10◦ / AUC@20◦ LO-RANSAC RANSAC SP+NN 15.7 / …
Figure 5
Figure 5. Figure 5: For the Bidirectional models, the MambaVision layer computes the following: yf = MambaVision(Norm(xin)). (13) yb = flip(MambaVision(flip(Norm(xin)))). (14) The outputs of Eqs. 13 and 14 are subsequently passed through the last linear layer, resulting in the final outpu…
Figure 6
Figure 6. Figure 6: Runtime Breakdown Extended. Flash Attention [5] enabled during inference time measurement on ScanNet [4]. its efficiency to prevent introducing additional learnable pa￾rameters, as the downsampling quality would depend on the learnt convolutional kernels. Area Interpol…
Figure 8
Figure 8. Figure 8: Rotation Invariance evaluation on the IMC 2020 [16] validation set. F. Runtime Breakdown extended VMatcher does not benefit from the usage of Flash Atten￾tion [5] compared to [3, 20, 30, 35, 41], due to the limited number of Transformer layers [39] in its architecture.…
Figure 7
Figure 7. Figure 7: VMatcher MambaVision layers scan directions visualisation. H. Limitations • The VMatcher model utilises the Mamba architecture, where a portion of the Mamba model is implemented in Triton [23], requiring a Graphics Processing Unit (GPU). While powerful hardware is comm…
Figure 9
Figure 9. Figure 9: Matching Examples on MegaDepth and ScanNet [4, 18] datasets. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 41 canonical work pages

  1. [1]

    Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors

    Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krys- tian Mikolajczyk. Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In CVPR, 2017. 5

    work page 2017

  2. [2]

    Surf: Speeded up robust features

    Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In Computer Vision – ECCV 2006, pages 404–417, Berlin, Heidelberg, 2006. Springer Berlin Heidelberg. 1, 2

    work page 2006

  3. [3]

    Aspanformer: Detector-free image matching with adaptive span transformer, 2022

    Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Ming- min Zhen, Tian Fang, David Mckinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer, 2022. 2, 3, 4, 5, 6, 7, 11

    work page 2022

  4. [4]

    Chang, Manolis Savva, Maciej Hal- ber, Thomas A

    Angela Dai, Angel X. Chang, Manolis Savva, Maciej Hal- ber, Thomas A. Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes.CoRR, abs/1702.04405, 2017. 6, 7, 8, 11, 12, 13

    arXiv 2017

  5. [5]

    Flashattention-2: Faster attention with better paral- lelism and work partitioning, 2023

    Tri Dao. Flashattention-2: Faster attention with better paral- lelism and work partitioning, 2023. 5, 11

    work page 2023

  6. [6]

    Transformers are ssms: General- ized models and efficient algorithms through structured state space duality, 2024

    Tri Dao and Albert Gu. Transformers are ssms: General- ized models and efficient algorithms through structured state space duality, 2024. 8

    work page 2024

  7. [7]

    Superpoint: Self-supervised interest point detection and description

    Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabi- novich. Superpoint: Self-supervised interest point detection and description. CoRR, abs/1712.07629, 2017. 2, 6, 7

    arXiv 2017

  8. [8]

    Repvgg: Making vgg-style convnets great again, 2021

    Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again, 2021. 3

    work page 2021

Show all 46 references
  1. [9]

    D2-net: A trainable cnn for joint description and detection of local features

    Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Polle- feys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint description and detection of local features. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 80...

    2019

  2. [10]

    Fischler and Robert C

    Martin A. Fischler and Robert C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6):381–395, 1981. 5

    1981

  3. [11]

    Silk: Sim- ple learned keypoints

    Pierre Gleize, Weiyao Wang, and Matt Feiszli. Silk: Sim- ple learned keypoints. In Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision (ICCV) , pages 22499–22508, 2023. 2, 5, 6

    2023

  4. [12]

    Mamba: Linear-time sequence mod- eling with selective state spaces, 2024

    Albert Gu and Tri Dao. Mamba: Linear-time sequence mod- eling with selective state spaces, 2024. 1, 2, 3, 4, 8, 11, 12

    2024

  5. [13]

    Combining recurrent, con- volutional, and continuous-time models with linear state- space layers, 2021

    Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher R ´e. Combining recurrent, con- volutional, and continuous-time models with linear state- space layers, 2021. 3 8

    2021

  6. [14]

    R. I. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision . Cambridge University Press, ISBN: 0521540518, second edition, 2004. 6

    2004

  7. [15]

    Mambavision: A hybrid mamba-transformer vision backbone, 2024

    Ali Hatamizadeh and Jan Kautz. Mambavision: A hybrid mamba-transformer vision backbone, 2024. 1, 2, 4, 10, 11

    2024

  8. [16]

    Image Matching across Wide Baselines: From Paper to Practice

    Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image Matching across Wide Baselines: From Paper to Practice. International Journal of Computer Vision, 2020. 11

    2020

  9. [17]

    PoseLib - Minimal Solvers for Camera Pose Estimation, 2020

    Viktor Larsson and contributors. PoseLib - Minimal Solvers for Camera Pose Estimation, 2020. 5

    2020

  10. [18]

    Megadepth: Learning single-view depth prediction from internet photos

    Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2041–2050, 2018. 6, 10, 12, 13

    2018

  11. [19]

    Jamba: A hybrid transformer-mamba language model, 2024

    Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, Omri Abend, Raz Alon, Tomer Asida, Amir Bergman, Roman Glozman, Michael Gokhman, Avashalom Manevich, Nir Ratner, Noam Rozen, Ere...

    2024

  12. [20]

    Lightglue: Local feature matching at light speed, 2023

    Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Polle- feys. Lightglue: Local feature matching at light speed, 2023. 2, 4, 5, 6, 7, 11

    2023

  13. [21]

    Vmamba: Visual state space model, 2024

    Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model, 2024. 2

    2024

  14. [22]

    David G. Lowe. Distinctive Image Features from Scale- Invariant Keypoints. International Journal of Computer Vi- sion, 60(2):91–110, 2004. 1, 2

    2004

  15. [23]

    Triton Inference Server: An Opti- mized Cloud and Edge Inferencing Solution., 2020

    NVIDIA Corporation. Triton Inference Server: An Opti- mized Cloud and Edge Inferencing Solution., 2020. 12

    2020

  16. [24]

    Yang, Zach DeVito, Mar- tin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zem- ing Lin, Natalia Gimelshein, Luca Antiga, Alban Desmai- son, Andreas K ¨opf, Edward Z. Yang, Zach DeVito, Mar- tin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Stei...

    1912 arXiv

  17. [25]

    Efficientvmamba: Atrous selective scan for light weight visual mamba, 2024

    Xiaohuan Pei, Tao Huang, and Chang Xu. Efficientvmamba: Atrous selective scan for light weight visual mamba, 2024. 2

    2024

  18. [26]

    Nascimento

    Guilherme Potje, Felipe Cadar, Andre Araujo, Renato Mar- tins, and Erickson R. Nascimento. Xfeat: Accelerated fea- tures for lightweight image matching, 2024. 2

    2024

  19. [27]

    R2d2: Repeatable and reliable detector and descrip- tor

    Jerome Revaud, Philippe Weinzaepfel, C ´esar De Souza, Noe Pion, Gabriela Csurka, Yohann Cabon, and Martin Humen- berger. R2d2: Repeatable and reliable detector and descrip- tor. arXiv preprint arXiv:1906.06195, 2019. 2

    1906 arXiv

  20. [28]

    Orb: An efficient alternative to sift or surf

    Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In 2011 International Conference on Computer Vision, pages 2564– 2571, 2011. 1, 2

    2011

  21. [29]

    From coarse to fine: Robust hierarchical localization at large scale, 2019

    Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale, 2019. 7

    2019

  22. [30]

    Superglue: Learning feature matching with graph neural networks, 2020

    Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks, 2020. 2, 5, 6, 7, 11

    2020

  23. [31]

    Benchmarking 6dof outdoor visual localiza- tion in changing conditions, 2018

    Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenborg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, Fredrik Kahl, and Tomas Pajdla. Benchmarking 6dof outdoor visual localiza- tion in changing conditions, 2018. 7

    2018

  24. [32]

    Structure-from-motion revisited

    Johannes Lutz Sch ¨onberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Com- puter Vision and Pattern Recognition (CVPR), 2016. 7, 10

    2016

  25. [33]

    Pixelwise view selection for un- structured multi-view stereo

    Johannes Lutz Sch ¨onberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for un- structured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 7, 10

    2016

  26. [34]

    Roformer: Enhanced transformer with rotary position embedding, 2023

    Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding, 2023. 4, 8

    2023

  27. [35]

    Loftr: Detector-free local feature matching with transformers

    Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In 2021 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 8918– 8927, 2021. 2, 3, 4, 5, 6, 7, 11

    2021

  28. [36]

    InLoc: Indoor visual localization with dense matching and view synthesis

    Hajime Taira, Masatoshi Okutomi, Torsten Sattler, Mircea Cimpoi, Marc Pollefeys, Josef Sivic, Tomas Pajdla, and Ak- ihiko Torii. InLoc: Indoor visual localization with dense matching and view synthesis. In CVPR, 2018. 7

    2018

  29. [37]

    Quadtree attention for vision transformers, 2022

    Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers, 2022. 2, 4, 6

    2022

  30. [38]

    Tyszkiewicz, Pascal Fua, and Eduard Trulls

    Michał J. Tyszkiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient, 2020. 2, 5, 6, 11

    2020

  31. [39]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need, 2023. 1, 2, 11

    2023

  32. [40]

    Matchformer: Interleaving attention in transformers for feature matching, 2022

    Qing Wang, Jiaming Zhang, Kailun Yang, Kunyu Peng, and Rainer Stiefelhagen. Matchformer: Interleaving attention in transformers for feature matching, 2022. 2, 3, 4

    2022

  33. [41]

    Efficient loftr: Semi-dense local feature matching with sparse-like speed, 2024

    Yifan Wang, Xingyi He, Sida Peng, Dongli Tan, and Xiaowei Zhou. Efficient loftr: Semi-dense local feature matching with sparse-like speed, 2024. 2, 3, 4, 5, 6, 7, 8, 11

    2024

  34. [42]

    Learning to find good correspondences

    Kwang Moo Yi, Eduard Trulls, Yuki Ono, Vincent Lepetit, Mathieu Salzmann, and Pascal Fua. Learning to find good correspondences. In 2018 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 2666–2674,

    2018

  35. [43]

    Xiaoming Zhao, Xingming Wu, Weihai Chen, Peter C. Y . Chen, Qingsong Xu, and Zhengguo Li. Aliked: A lighter keypoint and descriptor extraction network via deformable transformation, 2023. 6

    2023

  36. [44]

    Vision mamba: Efficient visual representation learning with bidirectional state space model, 2024

    Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model, 2024. 2, 10 9 Appendix A. VMatcher Model Configurations VMatcher configurations introduced in Sec. ...

    2024

  37. [45]

    was chosen as the variant for VMatcher’s architecture. E. Rotation Invariance DISK’s [38] rotation invariance evaluation was performed on the Image Matching Challenge (IMC) 2020 [16] valida- tion set. For each angle θ, 36 images are randomly selected and matched with their rot...

    2020

  38. [46]

    11 Figure 7

    Similarly, MambaVisions-Bi transposes the image, scans both the transposed image and its flipped version, then re- stores both to their original orientation. 11 Figure 7. VMatcher MambaVision layers scan directions visualisation. H. Limitations • The VMatcher model utilises th...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.