Pith. sign in

REVIEW 3 major objections 6 minor 47 references

CorrMoE: Mixture of Experts with De-stylization Learning for Cross-Scene and Cross-Domain Correspondence Pruning

T0 review · 3 major / 6 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 CorrMoE claims that de-stylizing graph features and routing them through a mixture of experts makes correspondence pruning generalize across scenes and domains.

desk verdict CorrMoE is a competent, incremental step for correspondence pruning; the in-domain numbers are plausible, but the cross-domain gains are not yet pinned on de-stylization. read the letter →

arxiv 2507.11834 v1 pith:WZAV77S2 submitted 2025-07-16 cs.CV

Peiwen Xia , Tangfei Liao , Wei Zhu , Danhuai Zhao , Jianjun Ke , Kaihao Zhang , Tong Lu , Tao Wang This is my paper · ORCID
classification cs.CV
keywords correspondencepruningtwo-viewgeometrymixtureofexpertsdomaingeneralizationMixStylede-stylizationcameraposeestimationoutlierrejection
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

This paper tries to establish that two-view correspondence pruning, which decides which tentative keypoint matches are geometrically consistent, can be made to generalize to unseen scenes and visual domains. It proposes CorrMoE, a framework that strips domain-specific style out of the learned graph features while routing those features through a mixture of small expert networks that specialize in different scene structures. The authors report that this combination outperforms prior state-of-the-art pruners on in-domain pose estimation and achieves a large gain on a 12-dataset zero-shot cross-domain benchmark. If the claim holds, off-the-shelf matchers could be deployed across driving, aerial, indoor, seasonal, and synthetic data without per-domain retraining.

What carries the argument

The load-bearing mechanism is the pairing of Progressive Mixstyle (PMix) with a Bi-Fusion Mixture-of-Experts module. PMix computes per-channel mean and standard deviation from a batch and its shuffled view, blends them with a Beta-distributed weight, and re-normalizes the features; unlike standard MixStyle the application probability increases linearly with training epoch, so early training focuses on in-domain patterns and later training injects simulated cross-domain styles. The Bi-Fusion MoE first fuses explicit and implicit graph features through FlowAttention, a linear-complexity attention, then a gating MLP produces top-k routing weights that select experts to modulate the fused feature. Together these modules carry the argument that domain shift is a style problem and scene diversity is a routing problem.

What would settle it

Hold the Bi-Fusion MoE fixed and replace PMix with plain MixStyle at a constant probability of 0.5, trained for the same 500k iterations and evaluated on the same 12 zero-shot datasets; if AUC@5 does not fall when the progressive schedule is removed, the ramp itself is not load-bearing. Similarly, replacing both PMix and the MoE with the BCLNet head should reproduce BCLNet's reported 14.42 mean AUC@5, confirming the comparison is operating on the same footing.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that correspondence pruning networks that fuse implicit and explicit local graphs can be made substantially more transferable by two coordinated changes: de-stylizing the graph features with a progressive MixStyle variant, and replacing the fixed fusion head with a Bi-Fusion Mixture-of-Experts layer. The De-stylization Dual Branch mixes channel-wise mean and standard deviation between shuffled batch samples, with the mixing probability rising from 0.2 to 0.5 during training, applied to both the implicit (order-aware pooling) and explicit (KNN attention) branches. The Bi-Fusion MoE then fuses the two branches with linear-complexity FlowAttention and routes the fused features through three experts per layer with top-1 selection. On YFCC100M this reaches 37.98 AUC@5, ahead of BCLNet's 35.70, and on the zero-shot benchmark it records 22.80 mean AUC@5 versus 14.42 for BCLNet, leading on most of the 12 datasets.

Load-bearing premise

The central premise is that mixing channel-wise means and standard deviations of graph features removes domain-specific style while leaving the geometric signal that tells inliers from outliers intact; if that mixing instead damages geometry, the reported cross-domain gains would not come from de-stylization.

Editorial extensions

If this is right

  • Correspondence pruning can be trained once on a single domain (YFCC100M) and still retain usable accuracy on driving, aerial, indoor, seasonal, and synthetic datasets, with mean AUC@5 about 8.4 points above the best prior method.
  • The explicit graph branch contributes more than the implicit branch when combined with the MoE, and the full dual-branch model outperforms either alone (37.98 vs 35.17 AUC@5 in the ablation).
  • Stacking the Bi-Fusion MoE four times gives the best accuracy-efficiency trade-off; three times is slightly worse and five times reduces both throughput and accuracy.
  • PMix outperforms both instance normalization and fixed-probability MixStyle on the in-domain YFCC100M test while matching MixStyle's cross-domain gain, suggesting de-stylization and in-domain accuracy are not in conflict.
  • Because expert selection is top-1 with only three experts per layer, the added scene-adaptation capacity comes without quadratic attention cost.

Reading between the lines

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

  • A testable extension implied but not run by the paper: fixing the MixStyle probability at its final value of 0.5 throughout training and comparing against the progressive schedule would isolate whether the ramp itself, rather than the mixing operation, drives the reported convergence and accuracy gains.
  • The de-stylization recipe is not tied to the specific dual-branch backbone; it could plausibly be dropped into other geometric learning tasks such as fundamental matrix estimation, homography filtering, or point cloud registration, where domain shift also corrupts putative correspondences.
  • Because the explicit branch carries more of the benefit in the ablation, a lightweight deployment variant that keeps the explicit branch and the MoE while dropping the implicit branch might preserve most of the cross-domain gain at lower compute.
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 / 6 minor

Summary. The paper proposes CorrMoE, a two-view correspondence pruning framework that combines a De-stylization Dual Branch with Progressive Mixstyle (PMix) and a Bi-Fusion Mixture-of-Experts module with linear-complexity attention and dynamic routing. The model is evaluated on in-domain datasets (YFCC100M, SUN3D), a cross-scene split of YFCC100M, and the zero-shot cross-domain ZEB benchmark, reporting state-of-the-art AUC values such as 37.98 vs 35.70 on YFCC100M and 22.80 vs 14.42 mean AUC@5 on ZEB.

Significance. If the reported results are reliable, CorrMoE would represent a notable improvement in cross-domain correspondence pruning, with a large absolute gain over BCLNet on ZEB. The paper contributes a novel application of MoE to this task and a schedule-adaptive MixStyle variant, and the authors have committed to releasing code. However, the core scientific claim that de-stylization drives the cross-domain gains is not supported by the current ablation design; the empirical evidence needs strengthening before the significance can be fully assessed.

major comments (3)
  1. [Section 4.4 (Table 8) and Section 4.3 (Table 5)] The central claim that de-stylization learning (PMix) is responsible for the cross-domain gains is not established. Table 8 reports cross-domain AUC@5 of 21.67 (IN), 22.39 (MixStyle), and 22.80 (PMix) when these modules are swapped inside the full model; the 0.41-point PMix-over-MixStyle difference is an order of magnitude smaller than the 8.38-point gap over BCLNet reported in Table 5. Because Table 6 ablates the Bi-Fusion MoE and the dual branches only on in-domain YFCC100M, there is no evidence that the large ZEB improvement is attributable to PMix rather than to the added MoE capacity, the learned routing gate, or the dual-branch architecture. The authors should report cross-domain ablations that remove the MoE and each branch, and include a matched-capacity control (e.g., BCLNet with the same MoE stack) on ZEB.
  2. [Section 4.1 and Tables 1-5] No uncertainty quantification is provided for any metric. Given the large variance in cross-domain performance (Table 5 shows dataset-level AUC@5 ranging from 6.45 to 36.33 for the same model) and the fact that the paper's headline claims are point estimates, the authors should report standard deviations over multiple training runs or at least per-dataset paired comparisons with significance tests (e.g., paired bootstrap or Wilcoxon) on the 12 ZEB datasets.
  3. [Section 4.1] Hyperparameter settings are tuned on the same YFCC100M benchmark used for final evaluation, and the manuscript does not state the validation split (the dataset description says 68 sequences for 'training and validation purposes' but no explicit split is given). This introduces a selection bias that can inflate in-domain scores; the paper should clarify the validation procedure and, ideally, fix hyperparameters before evaluating on the test subsets.
minor comments (6)
  1. [Section 4.1] The sentence 'pstart_mix and pstart_mix are set to 0.2 and 0.5' repeats the same variable; it should read pstart_mix and pend_mix.
  2. [Section 3.2 (Eq. 6)] The variables t and T are used without definition; clarify that t is the current epoch and T is the total number of training epochs, even though this is stated later in the text.
  3. [Section 3.3 (Eq. 25)] The symbol N is reused as the number of experts, which conflicts with N as the number of correspondences in Eq. (1); use n for the expert count.
  4. [Section 3.4 (Eq. 28)] The essential matrix loss is called Lcls in the text but should be Less; the subscripts [1] and [2] in the denominator are undefined.
  5. [Table 6] The row labels are ambiguous because checkmarks are placed without an explicit table grid; consider adding a 'Components' column with explicit check/cross entries for each row.
  6. [Section 4.3 (Table 4)] The paper splits the YFCC100M test set into four sub-scenes but does not state how the split is performed or confirm that these scenes are not in the training set; please clarify.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: central claims are evaluated on external benchmarks and supported by ablations; no derivation step reduces to its own inputs.

full rationale

The paper's load-bearing claims are empirical: CorrMoE reports superior pose-estimation AUC on the external YFCC100M, SUN3D, and Zero Shot Evaluation Benchmark, and the proposed modules are ablated against baselines. No quantity used as an input is also the predicted target. The PMix module (Eqs. 6-9) is defined as a progressive variant of MixStyle, but it is explicitly compared against MixStyle and Instance Normalization in Table 8 rather than assumed equivalent. The MoE and dual-branch components are validated by the Table 6 ablation, and hyperparameters such as MoE stack count are tested on the same benchmark, which is standard practice and does not make the result tautological. The paper includes self-citations to prior work by the same authors (e.g., VSFormer, CorrAdaptor), but these are used for architectural inspiration and are not load-bearing justifications for the cross-domain generalization claim. No equation is shown to be equivalent to another by construction, and no fitted parameter is relabeled as a prediction. The reported improvements may be attributable to multiple factors, but that is a question of experimental attribution, not circular reasoning.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

This is an empirical ML paper with no mathematical derivation. The central claim rests on benchmark evaluations and on design assumptions about how correspondence features encode domain style. All listed parameters are hand-set or tuned on the same benchmarks used for final reporting.

free parameters (6)
  • PMix start and end probabilities = 0.2 / 0.5
    Chosen by authors; control how quickly style mixing is introduced (Sec. 4.1).
  • Beta distribution alpha for MixStyle lambda = not specified
    Lambda ~ Beta(alpha, alpha) in Eqs. 7-8; alpha is never given, leaving a hidden tuning knob.
  • KNN graph neighbor count k = 9 (first), 6 (second pruning module)
    Hand-set in Sec. 4.1 and affects explicit graph construction.
  • OA Filter cluster count = 250
    Set in Sec. 4.1 for implicit graph pooling.
  • Number of Bi-Fusion MoE stacks and experts = 4 stacks, 3 experts, top-1
    Selected by ablation (Table 7); 4 stacks chosen as best trade-off.
  • Loss weight tau for essential matrix loss = not specified
    Eq. 26 defines Lcls + tau * Less, but tau is never stated.
assumptions (4)
  • domain assumption Channel-wise statistics encode domain style, and mixing them removes bias while preserving geometry
    Core premise of the PMix and De-stylization branch in Sec. 3.2.
  • domain assumption SIFT keypoints plus nearest-neighbor matching produce the correspondence set used by all compared methods
    Sec. 4.1; claims of state-of-the-art performance depend on this input pipeline.
  • domain assumption Evaluation on ZEB and on YFCC scene splits is a valid proxy for cross-domain and cross-scene generalization
    Sec. 4.3; the central empirical claim rests on these benchmarks.
  • standard math Weighted eight-point algorithm and full-size verification are correct and standard
    Eqs. 4-5, used to convert predicted inlier weights into pose estimates.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CorrMoE: Mixture of Experts with De-stylization Learning for Cross-Scene and Cross-Domain Correspondence Pruning." pith.science (2026). https://pith.science/paper/WZAV77S2

@misc{pith2026250711834,
  author       = {Pith},
  title        = {Pith review of: CorrMoE: Mixture of Experts with De-stylization Learning for Cross-Scene and Cross-Domain Correspondence Pruning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WZAV77S2}},
  note         = {Machine review of arXiv:2507.11834}
}
read the original abstract

Establishing reliable correspondences between image pairs is a fundamental task in computer vision, underpinning applications such as 3D reconstruction and visual localization. Although recent methods have made progress in pruning outliers from dense correspondence sets, they often hypothesize consistent visual domains and overlook the challenges posed by diverse scene structures. In this paper, we propose CorrMoE, a novel correspondence pruning framework that enhances robustness under cross-domain and cross-scene variations. To address domain shift, we introduce a De-stylization Dual Branch, performing style mixing on both implicit and explicit graph features to mitigate the adverse influence of domain-specific representations. For scene diversity, we design a Bi-Fusion Mixture of Experts module that adaptively integrates multi-perspective features through linear-complexity attention and dynamic expert routing. Extensive experiments on benchmark datasets demonstrate that CorrMoE achieves superior accuracy and generalization compared to state-of-the-art methods. The code and pre-trained models are available at https://github.com/peiwenxia/CorrMoE.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 39 canonical work pages

  1. [1]

    Agarwal, Y

    S. Agarwal, Y . Furukawa, N. Snavely, I. Simon, B. Curless, S. M. Seitz, and R. Szeliski. Building rome in a day. Communications of the ACM, 54(10):105–112, 2011

    2011

  2. [2]

    Barath, J

    D. Barath, J. Matas, and J. Noskova. Magsac: marginalizing sample consensus. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10197–10205, 2019

    work page 2019

  3. [3]

    Campos, R

    C. Campos, R. Elvira, J. J. G. Rodríguez, J. M. Montiel, and J. D. Tardós. Orb-slam3: An accurate open-source library for visual, visual– inertial, and multimap slam. IEEE Transactions on Robotics , 37(6): 1874–1890, 2021

    work page 2021

  4. [4]

    Chum and J

    O. Chum and J. Matas. Matching with prosac-progressive sample con- sensus. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 220–226. IEEE, 2005

    work page 2005

  5. [5]

    L. Dai, Y . Liu, J. Ma, L. Wei, T. Lai, C. Yang, and R. Chen. Ms2dg-net: Progressive correspondence learning via multiple sparse semantics dy- namic graph. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8973–8982, 2022

    work page 2022

  6. [6]

    L. Dai, X. Du, H. Zhang, and J. Tang. Mgnet: Learning correspon- dences via multiple graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3945–3953, 2024

    work page 2024

  7. [7]

    DeTone, T

    D. DeTone, T. Malisiewicz, and A. Rabinovich. Superpoint: Self- supervised interest point detection and description. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, pages 224–236, 2018

    work page 2018

  8. [8]

    Eigen, M

    D. Eigen, M. Ranzato, and I. Sutskever. Learning factored represen- tations in a deep mixture of experts. arXiv preprint arXiv:1312.4314, 2013

    arXiv 2013

Show all 47 references
  1. [9]

    Fedus, B

    W. Fedus, B. Zoph, and N. Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1–39, 2022

    2022

  2. [10]

    M. A. Fischler and R. C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated car- tography. Communications of the ACM, 24(6):381–395, 1981

    1981

  3. [11]

    Gross, M

    S. Gross, M. Ranzato, and A. Szlam. Hard mixtures of experts for large scale weakly supervised vision. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 6865–6873, 2017

    2017

  4. [12]

    M. I. Jordan and R. A. Jacobs. Hierarchical mixtures of experts and the em algorithm. Neural Computation, 6(2):181–214, 1994

    1994

  5. [13]

    Kerbl, G

    B. Kerbl, G. Kopanas, T. Leimkühler, and G. Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4), July 2023

    2023

  6. [14]

    D. P. Kingma and J. L. Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Represen- tations, pages 1–15, 2014

    2014

  7. [15]

    Lepikhin, H

    D. Lepikhin, H. Lee, Y . Xu, D. Chen, O. Firat, Y . Huang, M. Krikun, N. Shazeer, and Z. Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. arXiv preprint arXiv:2006.16668, 2020

    2006 arXiv

  8. [16]

    Z. Li, S. Zhang, and J. Ma. U-match: Exploring hierarchy-aware local context for two-view correspondence learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  9. [17]

    T. Liao, X. Zhang, Y . Xu, Z. Shi, and G. Xiao. Sga-net: A sparse graph attention network for two-view correspondence learning. IEEE Trans- actions on Circuits and Systems for Video Technology , 33(12):7578– 7590, 2023

    2023

  10. [18]

    T. Liao, X. Zhang, L. Zhao, T. Wang, and G. Xiao. Vsformer: Visual- spatial fusion transformer for correspondence pruning. In Proceedings of the AAAI Conference on Artificial Intelligence, 2024

    2024

  11. [19]

    Liu and J

    X. Liu and J. Yang. Progressive neighbor consistency mining for cor- respondence pruning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9527–9537, 2023

    2023

  12. [20]

    X. Liu, G. Xiao, R. Chen, and J. Ma. Pgfnet: Preference-guided filtering network for two-view correspondence learning. IEEE Transactions on Image Processing, 32:1367–1378, 2023

    2023

  13. [21]

    D. G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91–110, 2004

    2004

  14. [22]

    Y . Ma, Y . Zhuang, J. Hao, and I. King. 3d-moe: A mixture-of-experts multi-modal llm for 3d vision and pose diffusion via rectified flow. arXiv preprint arXiv:2501.16698, 2025

    2025 arXiv

  15. [23]

    Masoudnia and R

    S. Masoudnia and R. Ebrahimpour. Mixture of experts: a literature sur- vey. Artificial Intelligence Review, 42:275–293, 2014

    2014

  16. [24]

    X. Miao, G. Xiao, S. Wang, and J. Yu. Bclnet: Bilateral consensus learning for two-view correspondence pruning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4225–4232, 2024

    2024

  17. [25]

    Mur-Artal and J

    R. Mur-Artal and J. D. Tardós. Orb-slam2: An open-source slam sys- tem for monocular, stereo, and rgb-d cameras. IEEE Transactions on Robotics, 33(5):1255–1262, 2017

    2017

  18. [26]

    Mur-Artal, J

    R. Mur-Artal, J. M. M. Montiel, and J. D. Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE Transactions on Robotics, 31(5):1147–1163, 2015

    2015

  19. [27]

    Raguram, O

    R. Raguram, O. Chum, M. Pollefeys, J. Matas, and J.-M. Frahm. Usac: A universal framework for random sample consensus. IEEE Transac- tions on Pattern Analysis and Machine Intelligence , 35(8):2022–2038, 2012

    2022

  20. [28]

    Sarlin, A

    P.-E. Sarlin, A. Unagar, M. Larsson, H. Germain, C. Toft, V . Larsson, M. Pollefeys, V . Lepetit, L. Hammarstrand, F. Kahl, et al. Back to the feature: Learning robust camera localization from pixels to pose. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pa...

    2021

  21. [29]

    J. L. Schonberger and J.-M. Frahm. Structure-from-motion revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4104–4113, 2016

    2016

  22. [30]

    Shazeer, A

    N. Shazeer, A. Mirhoseini, K. Maziarz, A. Davis, Q. Le, G. Hinton, and J. Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017

    2017 arXiv

  23. [31]

    X. Shen, Z. Cai, W. Yin, M. Müller, Z. Li, K. Wang, X. Chen, and C. Wang. Gim: Learning generalizable image matcher from internet videos. In Proceedings of the International Conference on Learning Representations, 2024

    2024

  24. [32]

    Thomee, D

    B. Thomee, D. A. Shamma, G. Friedland, B. Elizalde, K. Ni, D. Poland, D. Borth, and L.-J. Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64–73, 2016

    2016

  25. [33]

    P. H. Torr and A. Zisserman. Mlesac: A new robust estimator with application to estimating image geometry. Computer Vision and Image Understanding, 78(1):138–156, 2000

    2000

  26. [34]

    S. Ullman. The interpretation of structure from motion. Proceedings of the Royal Society of London. Series B. Biological Sciences, 203(1153): 405–426, 1979

    1979

  27. [35]

    Ulyanov, A

    D. Ulyanov, A. Vedaldi, and V . Lempitsky. Instance normaliza- tion: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016

    2016 arXiv

  28. [36]

    H. Wu, J. Wu, J. Xu, J. Wang, and M. Long. Flowformer: Linearizing transformers with conservation flows. In Proceedings of the Interna- tional Conference on Machine Learning, 2022

    2022

  29. [37]

    J. Xiao, A. Owens, and A. Torralba. Sun3d: A database of big spaces reconstructed using sfm and object labels. In Proceedings of the IEEE International Conference on Computer Vision, pages 1625–1632, 2013

    2013

  30. [38]

    C. Yao, Y . Jia, H. Di, P. Li, and Y . Wu. A decomposition model for stereo matching. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 6091–6100, 2021

    2021

  31. [39]

    K. M. Yi, E. Trulls, Y . Ono, V . Lepetit, M. Salzmann, and P. Fua. Learn- ing to find good correspondences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2666– 2674, 2018

    2018

  32. [40]

    Zhang, D

    J. Zhang, D. Sun, Z. Luo, A. Yao, L. Zhou, T. Shen, Y . Chen, L. Quan, and H. Liao. Learning two-view correspondences and geometry using order-aware network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5845–5854, 2019

    2019

  33. [41]

    Zhang and J

    S. Zhang and J. Ma. Convmatch: Rethinking network design for two- view correspondence learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023

    2023

  34. [42]

    Zhang, Z

    S. Zhang, Z. Li, Y . Gao, and J. Ma. Dematch: Deep decomposition of motion field for two-view correspondence learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, pages 20278–20287, 2024

    2024

  35. [43]

    C. Zhao, Y . Ge, F. Zhu, R. Zhao, H. Li, and M. Salzmann. Progressive correspondence pruning by consensus learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 6464–6473, 2021

    2021

  36. [44]

    Zhong, G

    Z. Zhong, G. Xiao, L. Zheng, Y . Lu, and J. Ma. T-net: Effective permutation-equivariant network for two-view correspondence learn- ing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1950–1959, 2021

    1950

  37. [45]

    K. Zhou, Y . Yang, Y . Qiao, and T. Xiang. Domain generalization with mixstyle. arXiv preprint arXiv:2104.02008, 2021

    2021 arXiv

  38. [46]

    P. Zhu, Y . Sun, B. Cao, and Q. Hu. Task-customized mixture of adapters for general image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7099–7108, 2024

    2024

  39. [47]

    W. Zhu, Y . Liu, Y . He, T. Liao, K. Zheng, X. Xu, T. Wang, and T. Lu. Corradaptor: Adaptive local context learning for correspondence prun- ing. In Proceedings of the 27th European Conference on Artificial In- telligence, 2024

    2024

Pith tools

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