Pith. sign in

REVIEW 4 major objections 4 minor 53 references

Decoupled Geometric Parameterization and its Application in Deep Homography Estimation

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper proposes an eight-parameter geometric parameterization of homographies, split into similarity and kernel transformations, that lets deep networks estimate the homography by direct matrix multiplication instead of solving a…

desk verdict A useful parameterization paper with a real but fixable gap: the improved SKS factorization is asserted, not proven. read the letter →

arxiv 2505.16599 v2 pith:B5XNTRIJ submitted 2025-05-22 cs.CV

Yao Huang , Si-Yuan Cao , Yaqing Ding , Hao Yin , Shibin Xie , Shuting Wang , Zhijun Fang , Jiachun Wang
show 3 more authors
Shen Cai Junchi Yan Shuhan Shen
This is my paper · ORCID
classification cs.CV
keywords deephomographyestimationgeometricparameterizationsimilarity-kernel-similaritydecompositionangularoffsetspositionalprojectivetransformationdirectlineartransformmatrixmultiplication
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 aims to replace the standard four-corner positional-offset parameterization of homographies—used by nearly all deep homography networks—with a decoupled set of eight geometric parameters derived from a similarity-kernel-similarity factorization. The first four parameters describe a similarity transformation; the second four describe the kernel, the part that carries genuine projective distortion. The paper derives a linear relationship between the kernel parameters and four cotangent angular offsets, giving each parameter a concrete geometric meaning and explaining why a network can regress them. Because the homography matrix is reconstructed by multiplying the factor matrices together, the usual DLT linear-system solve disappears. Across three datasets and three network architectures, the proposed parameterization performs on par with or slightly better than four-corner offsets, and it also unifies similarity, affine, and homography estimation in one eight-parameter output.

What carries the argument

The load-bearing identity is the improved SKS factorization $H = H_{S1} H_{S2}^{-1} H_K H_{S2}$ (Eq. 11), with $H_{S2}$ a known normalization sending two correspondences to $[\mp 1, 0]^\top$, $H_{S1}$ of the form in Eq. (4), and $H_K$ of the form in Eq. (12). The linear relations in Eqs. (17) through (20) express the four kernel parameters as differences between cotangents of angles and $\cot 45^\circ$, making the kernel parameters visually accessible features. Together these identities convert homography estimation from solving an $8 \times 8$ linear system into estimating four similarity parameters and four kernel parameters.

What would settle it

Sample a random 3x3 homography, compute the four point correspondences it induces on a square, run the paper's normalization to get $H_{S2}$, solve for $H_{S1}$ and $H_K$ via the stated ordering in Eq. (11), and check whether the product $H_{S1}H_{S2}^{-1}H_KH_{S2}$ reproduces the original matrix exactly; a single mismatch for a valid homography disproves the factorization. Additionally, train the same network with angular-offset targets on randomly rotated or scaled versions of a dataset and compare convergence; if the kernel parameters cannot be learned directly, the empirical claim weakens.

Watch

Extended reading notes

Core claim

The central claim is that any planar homography can be decoupled into two independent sets of four geometric parameters via the improved SKS decomposition $H = H_{S1} H_{S2}^{-1} H_K H_{S2}$, where $H_{S1}$ is a similarity transformation carrying scale, rotation, and translation, and $H_K$ is the kernel transformation carrying the projective part. The paper further shows that the four kernel parameters are linearly related to the cotangents of four angles formed by the images of the canonical square under $H_K$, and that in the similarity stage the four parameters are linearly related to two-corner positional offsets. Consequently, homography estimation reduces to estimating two groups of four parameters that are explicit image features, and the homography matrix is obtained by matrix multiplication instead of solving a linear system.

Load-bearing premise

The reordered SKS factorization in Eq. (11) is stated without proof, and it changes the composition order of the original SKS theorem; if it fails for even one valid homography, the decoupling of the eight parameters, and with it the claim that the four kernel parameters can be estimated independently, collapses.

Editorial extensions

If this is right

  • Networks can emit the eight geometric parameters and reconstruct the homography as a product of 3x3 matrices, eliminating the DLT post-processing step from deep homography pipelines.
  • The four angular offsets provide a similarity-invariant measurement of projective distortion, so the same eight parameters can be compared across different similarity normalizations.
  • Degenerate cases are read off directly: $b_K = v_K = 0$ means an affine transformation, and $b_K = v_K = \Delta a_K = u_K = 0$ means a similarity transformation, so one network output covers all three transformation classes.
  • Across MSCOCO, SPID, and GoogleMaps, and across DHN, RHWF-2, and MCNet, the geometric parameterization matches or slightly improves the median angular error relative to four-corner positional offsets.

Reading between the lines

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

  • Not pursued in the paper: because the angular offsets are similarity-invariant, the split suggests a two-stage pipeline in which a network first estimates the similarity part and then the projective kernel separately, which could improve conditioning for large rotations or scales.
  • The linear relation between kernel parameters and cotangent angle offsets could be inverted to build a closed-form differentiable layer that maps predicted angles to $H_K$, making unsupervised training with angular losses straightforward.
  • A testable extension is to check whether angular-offset supervision on the kernel parameters helps specifically in strongly projective or cross-modal pairs, where positional-offset supervision is noisy.
  • The same decoupling might transfer to affine-core-affine or multi-plane homography estimation, a direction the authors themselves flag for future work.
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 an eight-degree-of-freedom homography parameterization based on a similarity-kernel-similarity (SKS) decomposition. It splits the eight parameters into two sets of four: a similarity transformation and a kernel transformation, and derives a linear relationship between the kernel parameters and four angular offsets (cotangents). This parameterization is then plugged into deep homography networks (DHN, RHWF, MCNet) in place of the standard four-corner positional offsets, allowing direct homography reconstruction by matrix multiplication rather than by solving a linear system. Experiments on MSCOCO, SPID, and GoogleMaps report that the geometric parameterization achieves accuracy comparable to four-corner offsets on the positional-offset metric and improved accuracy on the angular-offset metric.

Significance. If the claims hold, the paper gives a geometrically interpretable, end-to-end differentiable homography representation that is consistent with similarity and affine parameterizations. The angular-offset interpretation of the kernel parameters is a novelty that could generalize beyond homography estimation. The paper also ships a reproducibility statement with source code, and it evaluates the parameterization across multiple architectures and datasets. The main limitation is that the central factorization is asserted rather than proven, and the experimental comparisons are point estimates without uncertainty quantification. These issues are fixable and do not necessarily invalidate the approach.

major comments (4)
  1. [Sec. 3.2, Eq. (11)] The improved SKS factorization H = HS1 HS2^{-1} HK HS2 is asserted without proof, and its composition order differs from the original SKS theorem in Eq. (1). Since the angular-offset relations in Eqs. (17)-(20), the affine/similarity degeneracy criterion in Sec. 3.4, and the matrix-multiplication reconstruction all presuppose Eq. (11), the authors must supply a complete proof that every homography admits this form with HS2 a similarity normalizing {M,N} and HK of the specific 4-DOF form in Eq. (12). In particular, after factoring out the similarity HS1 determined by the two-point correspondence, the residual fixes M,N, and after conjugation by HS2 it fixes the canonical points; one must show that the 4-parameter family in Eq. (12) covers this residual exactly. The paper does not address this gauge issue.
  2. [Sec. 3.4, degeneracy criterion] The statement that a homography degenerates to an affine transformation if and only if bK = vK = 0, and to a similarity if and only if bK = vK = ΔaK = uK = 0, is given without derivation. The criterion is load-bearing for the claimed unified estimation of similarity and affine transformations in Sec. 4.5. Please provide a formal proof or an empirical validation, and specify the thresholds thresh1 and thresh2 introduced in Eqs. (B.28)-(B.29) of the Supplementary Material, including a sensitivity analysis.
  3. [Tables 1-2 and Figs. 6-7] All conclusions are based on median point estimates, and the reported improvements are often small (e.g., -0.4% to -8.6% in Table 2). There are no error bars, confidence intervals, or significance tests. Because the central claim is that the geometric parameterization performs 'comparable' to four-corner offsets, the absence of variance estimates makes it impossible to assess whether the observed differences are meaningful. Please report results over multiple training runs/seeds or, at minimum, bootstrap confidence intervals on the medians.
  4. [Sec. 4.5 and Supplementary C] The experiments on similarity and affine estimation using 'trained homography networks' do not specify whether the networks are evaluated zero-shot on the degenerate data or re-trained. This distinction matters for the 'unified estimation' claim: zero-shot transfer would be a strong result but needs a precise protocol, whereas re-training weakens the unification. The loss weighting between the four HS parameters and the four HK parameters (which have different scales) is also not described in the training procedure; please provide the exact loss function and any normalization used.
minor comments (4)
  1. [Eq. (14)] The notation is inconsistent: '∆ak' appears instead of '∆aK', and the origin point O is referenced without being defined in the surrounding text. Please clarify.
  2. [Fig. 3] The caption states that HS2 is 'known,' but the coordinate frames I2 and I3 are not defined explicitly before the figure. A one-sentence definition of each intermediate image would improve readability.
  3. [Conclusion] The claim that this is 'the first work to introduce angular offsets in vision tasks' is overbroad; angular offsets have appeared in pose estimation and in line-based homography methods. Please soften the claim or restrict it to the specific SKS setting.
  4. [Abstract / Introduction] The phrase 'eliminating the need for solving a linear system' is accurate, but the reconstruction still requires matrix multiplications with an inverse HS2^{-1}. The authors should state explicitly that HS2^{-1} has a closed-form simple expression so that the computational gain is transparent.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the derivations are coordinate transformations and the empirical claims rest on independent benchmarks.

full rationale

None of the claimed derivations reduces to its inputs by construction. In Sec. 3.1, HS is defined by Eq. (4) and the four corner offsets are derived in Eqs. (7)-(10) as an invertible linear map; this is a coordinate reparameterization, not a fitted result, and Sec. 4.2 compares independently trained networks on the same P.O. metric, so the 'on par' outcome is not forced. In Sec. 3.3, the coordinates of P2 and Q2 are computed in Eq. (15), and Eqs. (17)-(20) express the cotangent deviations as linear combinations of {ΔaK, bK, uK, vK}; because the linear map is invertible, the stated linear relation is a derived reparameterization, not a prediction derived from itself. The main proof gap is the 'improved SKS' factorization in Eq. (11), which is asserted rather than proved and relies on the SKS theorem [5] from an overlapping author group; this is a missing-proof / completeness concern, not circularity, since [5] states a mathematical decomposition whose assumptions do not include the present paper's target result, and the central experimental claims are validated against external datasets (MSCOCO, SPID, GoogleMaps) and baseline networks (DHN, RHWF, MCNet). No fitted parameter is renamed as a prediction: the ACE in A.O. of Table 1 is a derived metric computed from the same G.P. outputs, not an independent prediction claim. Therefore the paper does not exhibit circular reasoning.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The paper's central contribution rests on the SKS decomposition from prior work, on the unproved improved SKS form, and on the assumption that networks can regress angular offsets. Two thresholds for degenerate classification are chosen by hand and left unspecified. No new physical or mathematical entities are introduced.

free parameters (2)
  • thresh1 = unspecified
    Hand-chosen threshold in Eq. (B.28) to classify affine transformations in the unified estimation experiments; no value is reported in the paper.
  • thresh2 = unspecified
    Hand-chosen threshold in Eq. (B.29) to classify similarity transformations; no value is reported in the paper.
assumptions (3)
  • domain assumption SKS decomposition theorem from [5]: any homography can be written as H = H_{S2}^{-1} H_K H_{S1} with HS1, HS2 similarity and HK of the form in Eq. (2).
    Invoked in Sec. 2.3 as the starting point for the improved SKS; not reproved in this paper.
  • ad hoc to paper The improved SKS factorization H = H_{S1} H_{S2}^{-1} H_K H_{S2} (Eq. 11) is valid for all homographies.
    Stated in Sec. 3.2 without proof; the composition order differs from the original SKS theorem and its validity is not demonstrated.
  • domain assumption Neural networks can regress the proposed geometric parameters and angular offsets as effectively as pixel offsets.
    Relies on the empirical experiments and the universal approximation heuristic mentioned in the supplementary, not on a proven guarantee.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Decoupled Geometric Parameterization and its Application in Deep Homography Estimation." pith.science (2026). https://pith.science/paper/B5XNTRIJ

@misc{pith2026250516599,
  author       = {Pith},
  title        = {Pith review of: Decoupled Geometric Parameterization and its Application in Deep Homography Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B5XNTRIJ}},
  note         = {Machine review of arXiv:2505.16599}
}
read the original abstract

Planar homography, with eight degrees of freedom (DOFs), is fundamental in numerous computer vision tasks. While the positional offsets of four corners are widely adopted (especially in neural network predictions), this parameterization lacks geometric interpretability and typically requires solving a linear system to compute the homography matrix. This paper presents a novel geometric parameterization of homographies, leveraging the similarity-kernel-similarity (SKS) decomposition for projective transformations. Two independent sets of four geometric parameters are decoupled: one for a similarity transformation and the other for the kernel transformation. Additionally, the geometric interpretation linearly relating the four kernel transformation parameters to angular offsets is derived. Our proposed parameterization allows for direct homography estimation through matrix multiplication, eliminating the need for solving a linear system, and achieves performance comparable to the four-corner positional offsets in deep homography estimation.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 51 canonical work pages

  1. [1]

    Decoupled Geometric Parameterization and its Application in Deep Homography Estimation

    INTRODUCTION Planar homography, also known as two-dimensional (2D) projective transformation, is typically represented as a 3 ∗ 3 matrix. Due to homogeneous equality, a homography matrix contains 8 degrees of freedom (DOFs). Homography es- timation is essential for many computer vision tasks, such as camera calibration [45], pose estimation [11], image st...

    work page Pith review arXiv 2016

  2. [2]

    RELATED WORKS 2.1. Traditional Homography Estimation Traditional homography estimation often involves inter- est point extraction and matching, such as hand-crafted SIFT [27], SURF [4], ORB [33], and deep learning based LIFT [43], SOSNet [37], SuperPoint [13], and SuperGLUE [34]. Next, the DLT solver [17] is applied under the random sam- ple consensus (RA...

    work page

  3. [3]

    METHOD 3.1. Geometric Parameterization of Similarity Transfor- mation Before delving into the geometric parameterization of pro- jective transformations, we theoretically analyze why the geometric parameters estimation for similarity transforma- tions are accurate. Since similarity transformations have four DOF, they can be determined by two pairs of corr...

    work page

  4. [4]

    Configurations Datasets: We evaluate the homography estimation task across three commonly used datasets, including one dy- namic scene dataset and one cross-modal dataset

    EXPERIMENTS 4.1. Configurations Datasets: We evaluate the homography estimation task across three commonly used datasets, including one dy- namic scene dataset and one cross-modal dataset. A brief description of each dataset is provided below: • MSCOCO [23]: A large-scale image dataset widely used for evaluating homography estimation methods, consisted of...

    work page

  5. [5]

    CONCLUSION AND FUTURE WORK This paper presents a novel geometric parameterization of ho- mography that is suitable for estimation through neural net- works, based on the SKS decomposition. By introducing two independent sets of four geometric parameters, each with cor- responding projective distortion interpretations, our param- eterization not only align...

    work page

  6. [6]

    Extending DeepSDF for automatic 3D shape retrieval and similarity transform estimation

    O. Afolabi, A. Y . Yang, and S. S. Sastry. Extending deepsdf for automatic 3d shape retrieval and similarity transform estimation. arXiv preprint arXiv:2004.09048,

    work page Pith review arXiv 2004

  7. [7]

    Baker and I

    S. Baker and I. Matthews. Lucas-kanade 20 years on: A unifying framework. International Journal of Computer Vision, 56:221–255, 2004. 3

    work page 2004

  8. [8]

    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 (CVPR), 2019. 2, 6

    work page 2019

Show all 53 references
  1. [9]

    H. Bay, A. Ess, T. Tuytelaars, and L. Van Gool. Surf: Speeded up robust features. Computer Vision and Image Understanding (CVIU), 110(3):404–417, 2006. 2

    2006

  2. [10]

    S. Cai, Z. Wu, L. Guo, J. Wang, S. Zhang, J. Yan, and S. Shen. Fast and interpretable 2d homogra- phy decomposition: Similarity-kernel-similarity and affine-core-affine transformations. arXiv preprint arXiv:2402.18008, 2024. 2, 3, 5

    2024 arXiv

  3. [11]

    S.-Y . Cao, J. Hu, Z. Sheng, and H.-L. Shen. Iterative deep homography estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 6, 7

    2022

  4. [12]

    S.-Y . Cao, R. Zhang, L. Luo, B. Yu, Z. Sheng, J. Li, and H.-L. Shen. Recurrent homography estimation us- ing homography-guided image warping and focus trans- former. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,

  5. [13]

    Chang, C.-N

    C.-H. Chang, C.-N. Chou, and E. Y . Chang. Clkn: Cas- caded lucas-kanade networks for image alignment. In Proceedings of the IEEE Conference on Computer Vi- sion and Pattern Recognition (CVPR), 2017. 3, 6, 7, 13

    2017

  6. [14]

    Chee and Z

    E. Chee and Z. Wu. Airnet: Self-supervised affine reg- istration for 3d medical images using neural networks. arXiv preprint arXiv:1810.02583, 2018. 2

    2018 arXiv

  7. [15]

    X. Chen, Y . Meng, Y . Zhao, R. Williams, S. R. Vallabhaneni, and Y . Zheng. Learning unsupervised parameter-specific affine transformation for medical im- ages registration. In International Conference on Medi- cal Image Computing and Computer-Assisted Interven- tion(MICCAI), 2021. 2

    2021

  8. [16]

    Collins and A

    T. Collins and A. Bartoli. Infinitesimal plane-based pose estimation. International Journal of Computer Vision , 109:252–286, 2014. 1

    2014

  9. [17]

    DeTone, T

    D. DeTone, T. Malisiewicz, and A. Rabinovich. Deep image homography estimation. arXiv preprint arXiv: 1606.03798, 2016. 1, 2, 6, 7, 8, 13

    2016 arXiv

  10. [18]

    DeTone, T

    D. DeTone, T. Malisiewicz, and A. Rabinovich. Super- point: Self-supervised interest point detection and de- scription. In Conference on Computer Vision and Pat- tern Recognition Workshop (CVPRW) on Deep Learning for Visual SLAM, page 224–236, 2018. 2

    2018

  11. [19]

    Dosovitskiy, P

    A. Dosovitskiy, P. Fischer, E. Ilg, P. H ¨ausser, C. Hazir- bas, V . Golkov, P. van der Smagt, D. Cremers, and T. Brox. Flownet: Learning optical flow with convo- lutional networks. In Proceedings of the IEEE Interna- tional Conference on Computer Vision (ICCV), 2015. 2

    2015

  12. [20]

    M. A. Fischler and R. C. Bolles. Random sample con- sensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commu- nications of the ACM, 24(6):381–395, 1981. 1, 2, 6

    1981

  13. [21]

    A. P. Gee, D. Chekhlov, A. Calway, and W. Mayol- Cuevas. Discovering higher level structure in visual slam. IEEE Transactions on Robotics , 24(5):980–990,

  14. [22]

    Hartley and A

    R. Hartley and A. Zisserman. Multiple view geometry in computer vision, 2nd ed. Cambridge University Press,

  15. [23]

    M. Hong, Y . Lu, N. Ye, C. Lin, Q. Zhao, and S. Liu. Unsupervised homography estimation with coplanarity- aware gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,

  16. [24]

    Hornik, M

    K. Hornik, M. Stinchcombe, and H. White. Multi- layer feedforward networks are universal approxima- tors. Neural Networks, 1989. 13

    1989

  17. [25]

    H. Le, F. Liu, S. Zhang, and A. Agarwala. Deep homog- raphy estimation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), 2020. 1, 2, 6, 7

    2020

  18. [26]

    Lee and J.-Y

    K.-Y . Lee and J.-Y . Sim. Warping residual based im- age stitching for large parallax. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1

    2020

  19. [27]

    J. Lin, Z. Wei, C. Ding, and K. Jia. Category-level 6d object pose and size estimation using self-supervised deep prior deformation networks. In Proceedings of the European Conference on Computer Vision (ECCV),

  20. [28]

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll ´ar, and C. L. Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision (ECCV),

  21. [29]

    X. Lin, W. Yang, Y . Gao, and T. Zhang. Instance- adaptive and geometric-aware keypoint learning for category-level 6d object pose estimation. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2

    2024

  22. [30]

    S. Liu, Y . Lu, H. Jiang, N. Ye, C. Wang, and B. Zeng. Unsupervised global and local homography estimation with motion basis learning. IEEE Transactions on Pat- tern Analysis and Machine Intelligence , 45(6):7885– 7899, 2023. 1, 3

    2023

  23. [31]

    S. Liu, N. Ye, C. Wang, J. Zhang, L. Jia, K. Luo, J. Wang, and J. Sun. Content-aware unsupervised deep homography estimation and its extensions. IEEE Trans- actions on Pattern Analysis and Machine Intelligence , 45(3):2849–2863, 2023. 2

    2023

  24. [32]

    D. G. Lowe. Distinctive image features from scale- invariant keypoints. International Journal of Computer Vision, 60:91–110, 2004. 2, 6

    2004

  25. [33]

    T. C. W. Mok and A. C. S. Chung. Affine medical image registration with coarse-to-fine vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2

    2022

  26. [34]

    Mur-Artal, J

    R. Mur-Artal, J. M. M. Montiel, and J. D. Tard ´os. Orb- slam: A versatile and accurate monocular slam sys- tem. IEEE Transactions on Robotics, 31(5):1147–1163,

  27. [35]

    Nguyen, S

    T. Nguyen, S. W. Chen, S. S. Shivakumar, C. J. Tay- lor, and V . Kumar. Unsupervised deep homogra- phy: A fast and robust homography estimation model. IEEE Robotics and Automation Letters , 3(3):2346– 2353, 2018. 1, 2, 6

    2018

  28. [36]

    F. E. Nowruzi, R. Laganiere, and N. Japkowicz. Ho- mography estimation from image pairs with hierarchical convolutional networks. In Proceedings of the IEEE In- ternational Conference on Computer Vision Workshops (ICCVW), 2017. 1, 6, 7

    2017

  29. [37]

    Raguram, O

    R. Raguram, O. Chum, M. Pollefeys, and et al. Usac: A universal framework for random sample consensus. IEEE Transactions on Pattern Analysis and Machine In- telligence, 35(8):2022–2038, 2013. 2

    2022

  30. [38]

    Rublee, V

    E. Rublee, V . Rabaud, K. Konolige, and G. Bradski. Orb: an efficient alternative to sift or surf. In Inter- national Conference on Computer Vision (ICCV), pages 2564–2571, 2011. 2

    2011

  31. [39]

    Sarlin, D

    P.-E. Sarlin, D. DeTone, T. Malisiewicz, and A. Ra- binovich. Superglue: Learning feature matching with graph neural networks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2020. 2

    2020

  32. [40]

    R. Shao, G. Wu, Y . Zhou, Y . Fu, L. Fang, and Y . Liu. Localtrans: A multiscale local transformer network for cross-resolution homography estimation. In Pro- ceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 7

    2021

  33. [41]

    D. Sun, X. Yang, M.-Y . Liu, and J. Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) ,

  34. [42]

    Y . Tian, X. Yu, B. Fan, F. Wu, H. Heijnen, and V . Bal- ntas. Sosnet: Second order similarity regularization for local descriptor learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 11008–11017, 2019. 2

    2019

  35. [43]

    C. Wang, D. Xu, Y . Zhu, R. Mart ´ın-Mart´ın, C. Lu, L. Fei-Fei, and S. Savarese. Densefusion: 6d object pose estimation by iterative dense fusion. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2

    2019

  36. [44]

    D. Wang, C. Zhang, H. Cheng, Y . Shang, and L. Mei. Spid: Surveillance pedestrian image dataset and perfor- mance evaluation for pedestrian detection. In Proceed- ings of the Asian Conference on Computer Vision Work- shops (ACCVW), 2017. 6, 13

    2017

  37. [45]

    E. W. Weisstein. Diamond. https://mathworld. wolfram.com/Diamond.html, 2003. 4

    2003

  38. [46]

    Xiang, T

    Y . Xiang, T. Schmidt, V . Narayanan, and D. Fox. Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes. In Proceedings of Robotics: Science and Systems (RSS), 2018. 2

    2018

  39. [47]

    N. Ye, C. Wang, H. Fan, and S. Liu. Motion basis learn- ing for unsupervised deep homography estimation with subspace projection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) ,

  40. [48]

    K. M. Yi, E. Trulls, V . Lepetit, and P. Fua. Lift: Learned invariant feature transform. In European Conference on Computer Vision (ECCV), pages 467–483, 2016. 2

    2016

  41. [49]

    Zhang, C

    J. Zhang, C. Wang, S. Liu, L. Jia, N. Ye, J. Wang, J. Zhou, and J. Sun. Content-aware unsupervised deep homography estimation. In Proceedings of the Euro- pean Conference on Computer Vision (ECCV), 2020. 1, 2, 6

    2020

  42. [50]

    Z. Zhang. A flexible new technique for camera calibra- tion. IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 22(11):1330–1334, 2000. 1

    2000

  43. [51]

    Y . Zhao, X. Huang, and Z. Zhang. Deep lucas-kanade homography for multimodal image alignment. In Pro- ceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), 2021. 3, 7

    2021

  44. [52]

    Q. Zhou, T. Sattler, and L. Leal-Taix ´e. Patch2pix: Epipolar-guided pixel-level correspondences. In Pro- ceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), 2021. 2

    2021

  45. [53]

    Zhu, S.-Y

    H. Zhu, S.-Y . Cao, J. Hu, S. Zuo, B. Yu, J. Ying, J. Li, and H.-L. Shen. Mcnet: Rethinking the core ingredients for accurate and efficient homography estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 6, 7, 8, ...

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.