Pith. sign in

REVIEW 2 major objections 5 minor 68 references

Registration beyond Points: General Affine Subspace Alignment via Geodesic Distance on Grassmann Manifold

T0 review · 2 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 The paper claims to be the first to derive an optimizable cost function for aligning affine subspaces—lines and planes—by minimizing geodesic distance on the Grassmann manifold directly as a function of rotation $\mathbf{R}$ and…

desk verdict A useful new cost function for affine subspace registration, but the translation BnB's global optimality claim rests on an unproven, likely invalid bound. read the letter →

arxiv 2507.17998 v2 pith:4RDLPOTB submitted 2025-07-24 cs.CV

Jaeho Shin , Hyeonjae Gil , Junwoo Jang , Maani Ghaffari , Ayoung Kim This is my paper · ORCID
classification cs.CV
keywords affineGrassmannianGrassmannmanifoldsubspaceregistrationgeodesicdistancebranchandboundinlier-setmaximizationline-to-planeRGB-Dodometry
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

Existing line and plane registration methods either approximate features with points or parameterize them in Euclidean space, where sign ambiguity (a plane normal versus its negative) distorts the metric. This paper instead works on the affine Grassmannian, the manifold of affine subspaces, and claims the first cost function that expresses the geodesic distance between two Grassmannian features explicitly in terms of a rigid transformation $(\mathbf{R}, \mathbf{t})$. The key move is a proof that an embedded subspace's orthonormal basis gives an exact representation of the cost: a smaller subspace lies at zero geodesic distance from a larger one exactly when each of its basis vectors is spanned by the larger subspace's basis. The resulting cost is optimizable and representation-agnostic, and a branch-and-bound extension finds globally optimal inlier sets for 3D line-to-line, line-to-plane, and plane-to-plane registration. The authors demonstrate improved accuracy on object registration, RGB-D odometry, perspective-n-line pose estimation, and correspondence-free localization.

What carries the argument

The central object is the affine Grassmannian embedding, the map from an affine subspace to a linear subspace one dimension higher, together with the basis-spans-subspace criterion from Theorem 2. The embedding lets a line or plane be represented by an orthonormal matrix whose columns are the transformed source basis and a normalized displacement vector; the Grassmann distance between two such matrices is the metric the paper wants to optimize. The basis criterion replaces the non-differentiable singular-value computation of principal angles with squared Euclidean residuals between each basis vector and its projection onto the transformed target subspace. These residuals are polynomial in $\mathbf{R}$ and $\mathbf{t}$, so the cost can be minimized by ordinary nonlinear least squares and bounded for branch-and-bound search.

What would settle it

Generate random line correspondences and, for a fixed rotation and a small translation cube, compute the true maximum over $\mathbf{t}$ of $\| (\tilde{b}_2^\top \tilde{b}_1'(R^*, t_0)) \tilde{b}_1'(R^*, t_0) - (\tilde{b}_2^\top \tilde{b}_1'(R^*, t)) \tilde{b}_1'(R^*, t) \|$ using dense sampling or local optimization, and compare it with the maximum over the cube's vertices; finding any interior point with a larger value than all vertices would invalidate Eq. (62), making the lower bound unsound and allowing the branch-and-bound search to prune the global optimum.

Watch

Extended reading notes

Core claim

Using the orthonormal matrix representation of the affine Grassmannian—embedding a $k$-dimensional affine subspace $A + b_0$ into $\mathrm{Gr}(k+1, n+1)$ via $Y_{z(A+b_0)} = \begin{pmatrix} A & b_0/\sqrt{1+\|b_0\|^2} \\ 0 & 1/\sqrt{1+\|b_0\|^2} \end{pmatrix}$—the paper proves that the $\mathrm{SE}(n)$ group action moves the subspace's linear part and displacement in a closed form (Theorem 1), and that zero Grassmann distance between two embedded subspaces of possibly different dimension is equivalent to every orthonormal basis vector of the smaller one being in the span of the larger one (Theorem 2). From this it constructs Problem 2, Eq. (7): for $N$ paired affine primitives, minimize $\sum_i \left( \sum_{j=1}^{k_i} \| P_{R \cdot B_i} a_i^j - a_i^j \|_2^2 + \| P_{z(T \cdot (B_i + d_i^0))} \tilde{c}_i^0 - \tilde{c}_i^0 \|_2^2 \right)$, where $P$ denotes orthogonal projection onto the transformed source subspace. The paper claims that the zero set of this cost coincides exactly with zero geodesic distance, that the cost separates into rotation-only and translation-dependent terms, and that exact bounds can be derived for a branch-and-bound inlier-set maximizer. Experiments across four computer vision tasks support the claim that this cost improves convergence of existing linear solvers or outperforms them.

Load-bearing premise

The translation branch-and-bound lower bound rests on the claim that, along the arc traced by the normalized displacement as $\mathbf{t}$ varies within a cube, the inner product with the fixed embedded target vector is always monotonic, convex, or concave, so the residual difference of Eq. (60) is maximized at a cube vertex; the paper supports this only empirically, writing "in every tested case in our experiments," and gives no proof.

Editorial extensions

If this is right

  • Rotation and translation can be estimated in stages: the linear-subspace terms in Eq. (7) depend only on $\mathbf{R}$, so an inlier-maximizing rotation search can be run first and the translation search second.
  • The same cost applies to any pair of affine subspaces of different dimensions, including lines-to-planes, which lets the perspective-n-line problem be reformulated as a line-to-plane registration.
  • With the derived bounds, the branch-and-bound solver finds a deterministic global optimum of the inlier set, effective at outlier ratios up to 80% in synthetic PnL experiments, and supports correspondence-free localization.
  • Because the cost is agnostic to the sign of basis vectors, refining a conventional parameter-based linear solver with this cost reduces translation error, as shown in object registration and RGB-D odometry.

Reading between the lines

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

  • The basis-spanning criterion is metric-free, so the cost could be generalized to weighted or robustified residuals, folding in feature covariance or correspondence confidence without changing the zero-distance equivalence; the paper does not develop this.
  • The appendix's numerical observation that the two antipodal straight-line parameter paths map to arcs whose lengths sum to $\pi$, with one matching the geodesic to numerical precision, hints at an explicit closed-geodesic formula on $\mathrm{Gr}(2,3)$; if pursued, it could yield a direct differentiable expression for the longer arc connecting two subspaces.
  • The cost is defined for arbitrary dimensions $n$ and $k$, so it should extend beyond 3D lines and planes to higher-dimensional affine subspace features in structure-from-motion or model fitting; this is a direct generalization of the presented framework.
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

2 major / 5 minor

Summary. The paper proposes a registration cost for affine subspaces (lines and planes) by embedding them into a higher-dimensional Grassmannian and measuring projection residuals of the embedded bases. It derives an SE(n) action on the affine Grassmannian (Theorem 1), proves that zero Grassmann distance is equivalent to zero projection residual (Theorem 2), and formulates a cost function explicitly parameterized by rotation R and translation t (Problem 2, Eq. (7)). The authors then develop a branch-and-bound solver that maximizes inlier sets for rotation and translation and evaluate the method on object registration, RGB-D odometry, perspective-n-line, and correspondence-free localization tasks.

Significance. If the cost function and the bounds were fully validated, the paper would provide a unified geometric framework for line/plane registration that avoids the sign ambiguity of vector parameterizations. The authors give self-contained derivations of the group action and of the zero-set equivalence, release code, and compare against standard baselines on public datasets. The two-stage BnB inlier-set maximization pipeline is a practical contribution. However, the advertised equivalence with geodesic-distance minimization and the deterministic global-optimality guarantee rest on points that need correction or additional proof, so the paper requires major revision.

major comments (2)
  1. [§4.2, Eq. (7); Appendix 7.4] The abstract and Section 4 claim that Problem 2 (Eq. (7)) minimizes the squared Grassmann geodesic distance of Eq. (6). Appendix 7.4 proves only that the residual in Eq. (31) is zero if and only if the Grassmann distance is zero; it does not establish equality between the summed squared projection residuals in Eq. (7) and the summed squared geodesic distances in Eq. (6). For 1D subspaces, the projection residual is |sin θ| while the geodesic distance is θ, so the two objectives differ away from zero. The phrase 'equivalently optimize Eq. (6)' is therefore unsupported, and the claim of 'minimizing the geodesic distance' should be replaced by a statement that Eq. (7) is a chordal/projection cost with the same zero set, unless a new proof of cost equivalence is supplied.
  2. [Appendix 8.4.1, Eqs. (60)–(63)] The translation lower bound relies on the assertion that Eq. (60) is maximized at a vertex of the translation cube because the inner products (tilde_b2)^T tilde_b1'(R*, t) are 'monotonic increasing, monotonic decreasing, convex, or concave', justified only by 'in every tested case in our experiments'. This is not a proof, and the geometric statement preceding it is inaccurate: t -> R*b1 + R*(I - dd^T)R*^T t maps the cube through a rank-2 linear map, so after normalization the image is a 2D surface in S^3, not a line segment with two endpoints. A concave inner-product profile can place the maximum of Eq. (60) in the interior of the cube, in which case Eq. (62) overestimates the lower bound and the BnB in Algorithm 5 can prune the global optimum. Because this bound is the only support for the deterministic global-optimality claims in the abstract and Section 5, the authors must either prove a vertex-maximum property or replace the bound with a valid one.
minor comments (5)
  1. [Appendix 8.4.1] The text references 'Theorem 3' without defining it; presumably Theorem 1 or Corollary 1.1 is intended.
  2. [Algorithms 3–5] The condition 'If.size < ¯νr' compares an integer inlier count with a real-valued bound; please clarify the intended comparison and notation.
  3. [§4.2, Eq. (8)–(9)] Because the method solves Eq. (8) for R and then Eq. (9) for t, the manuscript should state explicitly that the two-stage procedure does not guarantee joint global optimality of Problem 2 unless the cost separates exactly.
  4. [Appendix 9.2, Fig. 9–10] The empirical observations that the projected curve length equals the geodesic length and that l1 + l2 = π are reported without proof; either add a derivation or soften the claim.
  5. [§5.4, Correspondence-free Localization] The localization experiment reports 0.65° and 0.03% error but gives no number of test frames and no comparison baseline; please add these details.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the cost derivation is self-contained and the flagged translation-bound gap is an unproven assumption, not a circular step.

full rationale

The manuscript constructs its central cost, Problem 2 / Eq. (7), from Theorem 1 (the SE(n) group action on the affine Grassmannian) and Theorem 2 (zero Grassmann distance iff every basis of the smaller subspace is spanned by the larger subspace), both proved in the appendix from standard Grassmann geometry. No parameter is fitted to data and then renamed as a prediction; the experiments compare against external baselines and public datasets. The only load-bearing concern raised by the reviewer, the translation BnB bound in Appendix 8.4.1, rests on the empirical assertion that inner-product profiles are monotonic/convex/concave 'in every tested case in our experiments,' which is an unproven correctness assumption about the bound, not a circular reduction of the result to its own inputs. The paper's citations to prior work, including the Go-ICP rotation bound, are external standard lemmas rather than self-citations that smuggle in the desired conclusion. Finally, while the projection-residual cost in Eq. (7) is not numerically equal to the squared geodesic distance away from the zero set, the paper only proves and uses zero-set equivalence for exact alignment, so this is a fidelity or correctness issue rather than circularity. No enumerated circularity pattern is present.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The derivation relies on standard manifold geometry, the affine Grassmannian embedding, and one unproven empirical assumption about the shape of the objective function needed for the BnB bound.

free parameters (1)
  • Inlier thresholds epsilon_R, epsilon_t = Not stated in paper; set per experiment
    Used in the BnB inlier-set maximization and termination criteria; hand-chosen thresholds affect which correspondences count as inliers.
assumptions (4)
  • standard math Standard geometry of the Grassmann manifold, including principal angles and geodesic distance via SVD.
    Used to define d_Gr and to justify Theorem 2 in Appendix 8.1.
  • domain assumption The z-embedding of affine subspaces into Gr(k+1, n+1) preserves the relevant metric geometry.
    Problem 2 is built on this embedding, Definition 3.
  • standard math Triangle inequality and the Go-ICP bound d_Gr(R0 d, R d) <= min(pi/2, sqrt(3) sigma_r) hold for the rotation search cubes.
    Used in Appendix 8.3 to derive BnB bounds, taken from reference [40].
  • ad hoc to paper The function in Eq. (60) is maximized at a cube vertex because the inner products are monotonic or convex/concave.
    The translation BnB lower bound Eq. (63) depends on this; only empirically checked, not proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Registration beyond Points: General Affine Subspace Alignment via Geodesic Distance on Grassmann Manifold." pith.science (2026). https://pith.science/paper/4RDLPOTB

@misc{pith2026250717998,
  author       = {Pith},
  title        = {Pith review of: Registration beyond Points: General Affine Subspace Alignment via Geodesic Distance on Grassmann Manifold},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4RDLPOTB}},
  note         = {Machine review of arXiv:2507.17998}
}
abstract

Affine Grassmannian has been favored for expressing proximity between lines and planes due to its theoretical exactness in measuring distances among features. Despite this advantage, the existing method can only measure the proximity without yielding the distance as an explicit function of rigid body transformation. Thus, an optimizable distance function on the manifold has remained underdeveloped, stifling its application in registration problems. This paper is the first to explicitly derive an optimizable cost function between two Grassmannian features with respect to rigid body transformation ($\mathbf{R}$ and $\mathbf{t}$). Specifically, we present a rigorous mathematical proof demonstrating that the bases of high-dimensional linear subspaces can serve as an explicit representation of the cost. Finally, we propose an optimizable cost function based on the transformed bases that can be applied to the registration problem of any affine subspace. Compared to vector parameter-based approaches, our method is able to find a globally optimal solution by directly minimizing the geodesic distance which is agnostic to representation ambiguity. The resulting cost function and its extension to the inlier-set maximizing Branch-and-Bound (BnB) solver have been demonstrated to improve the convergence of existing solutions or outperform them in various computer vision tasks. The code is available on https://github.com/joomeok/GrassmannRegistration.

Figures

Figures reproduced from arXiv: 2507.17998 by the authors.

Figure 1
Figure 1. Comparison of parameter and geodesic-based reg￾istration. Existing methods for solving line and plane registra￾tion minimize the squared sum of distances between vectors used to parameterize features in Euclidean space (e.g., Plucker coordi- ¨ nates, plane parameters). However, the sign ambiguity of vector representations induces metric inconsistencies and suboptimality. In contrast, as geodesic distances represente… view at source ↗
Figure 2
Figure 2. Registration of two 2D lines by separate estimation. (Left) Given a pair of 2D lines (l1, l2), our inlier-set maximization algorithm estimates the rotation matrix by making each basis of linear subspace aligned, which is a 2D direction vector. (Middle, Right) After obtaining the rotation, a translation vector is estimated by inducing the secondary basis vector of the embedded subspace z(l2), denoted as b˜2, to be in… view at source ↗
Figure 3
Figure 3. Space Station data registration results. The point￾based method (Olsson) yields a severe performance drop when es￾timating rotation due to its vulnerability to noise. Parameter-based (LinEq) alignment fails more critically in estimating translation because explicitly selecting one of two antipodal parameters dis￾torts the translation cost. lational error. The error from Olsson’s method steadily in￾creases with respe… view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Estimated trajectories of Replica dataset. For both Office3 (left) and Room0 (right), all other benchmarks exhibit sig￾nificant misalignment with the ground truth trajectory, in contrast to the high accuracy achieved by our approaches. and uncentered cases, with the ce…
Figure 7
Figure 7. Figure 7: Our work can also align partially overlapped line recon [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 5
Figure 5. Figure 5: PnL experiment results with synthetic data for cen￾tered (top) and uncentered (bottom) cases. Our algorithm ro￾bustly estimates rotation and translation by solving the inlier set maximization problem, even under the outlier ratio of 80%. 0 50 100 150 100 80 60 40 20 0 …
Figure 6
Figure 6. Figure 6: Additional results from the synthetic PnL experi￾ment. (a) Convergence of the bounds on the inliers for 100 corre￾spondences with a 10% outlier ratio. (b) Runtime with respect to endpoint pixel noise and the number of correspondences. scenarios, even under a severe out…
Figure 8
Figure 8. Figure 8: Measurement points on the model with outlier p. This figure illustrates measurement points without noise (blue) lying on the corresponding model lines. When an outlier p (red), expressed by parameters α and β, is paired with line l, the optimal rotation R(θ) may adjust…
Figure 9
Figure 9. Figure 9: Comparison of two projected straight lines in parameter space on the manifold. This figure compares two curves connecting the initial line v1 = (1, 2, −5) and the end line v2 = (1, −3, 5). The left figure illustrates the projected trajectory of the straight line connec…
Figure 10
Figure 10. Figure 10: Mapped straight lines on manifold. Two distinct straight lines connecting the features, differing only by the sign of one parameter, may form a closed geodesic on the Grassmann manifold. Additionally, the projection of the shorter line (blue) may align with the geodes…
Figure 11
Figure 11. Figure 11: Elapsed time analysis on object registration and RGB-D odometry tasks. (a) We examine the effect of the outlier ratio on the computational time for the object registration task. We show a comparable computational speed to the approximated parameter-based method while …
Figure 12
Figure 12. Figure 12: Elapsed time of PnL experiment with synthetic data. PnL task. As shown in [PITH_FULL_IMAGE:figures/full_fig_p023_12.png]
Figure 13
Figure 13. Figure 13: Failure case of Olsson [29] in Room1 sequence. This section presents visualizations of the frame pairs where each algorithm in the RGB-D odometry experiment recorded the largest error and includes a brief analysis of the causes. As shown in [PITH_FULL_IMAGE:figures/f…
Figure 14
Figure 14. Figure 14: Failure case of Park [31] in Office3 sequence . Park generalizes the discrete intensity function obtained from the image into a continuous representation in 3D space by utilizing the gradient at a specific point. The gradient is estimated by minimizing the difference …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

68 extracted references · 66 canonical work pages

  1. [1]

    Ceres solver, 2022

    Sameer Agarwal, Keir Mierle, and The Ceres Solver Team. Ceres solver, 2022. 6

    work page 2022

  2. [2]

    CvxP- nPL: A unified convex solution to the absolute pose estima- tion problem from point and line correspondences

    S ´ergio Agostinho, Jo˜ao Gomes, and Alessio Del Bue. CvxP- nPL: A unified convex solution to the absolute pose estima- tion problem from point and line correspondences. Journal of Mathematical Imaging and Vision, 65(3):492–512, 2023. 7, 8, 13

    work page 2023

  3. [3]

    Geometric mean and geodesic regression on grass- mannians

    Ekkehard Batzies, Knut H ¨uper, Luis Machado, and F Silva Leite. Geometric mean and geodesic regression on grass- mannians. Linear Algebra and its Applications, 466:83–101,

    work page

  4. [4]

    A grassmann manifold handbook: Basic geometry and compu- tational aspects

    Thomas Bendokat, Ralf Zimmermann, and P-A Absil. A grassmann manifold handbook: Basic geometry and compu- tational aspects. Advances in Computational Mathematics , 50(1):6, 2024. 10

    work page 2024

  5. [5]

    Numerical methods for computing angles between linear subspaces

    ¯Ake Bj ¨orck and Gene H Golub. Numerical methods for computing angles between linear subspaces. Mathematics of computation, 27(123):579–594, 1973. 4, 3

    work page 1973

  6. [6]

    Convex global 3D registration with Lagrangian duality

    Jesus Briales and Javier Gonzalez-Jimenez. Convex global 3D registration with Lagrangian duality. InProc. IEEE Conf. on Comput. Vision and Pattern Recog. , pages 4960–4969,

    work page

  7. [7]

    Globally optimal 2D-3D registration from points or lines without correspondences

    Mark Brown, David Windridge, and Jean-Yves Guillemaut. Globally optimal 2D-3D registration from points or lines without correspondences. In Proc. IEEE Intl. Conf. on Com- put. Vision, pages 2111–2119, 2015. 4

    work page 2015

  8. [8]

    Globally-optimal inlier set maximisation for camera pose and correspondence estimation

    Dylan Campbell, Lars Petersson, Laurent Kneip, and Hong- dong Li. Globally-optimal inlier set maximisation for camera pose and correspondence estimation. IEEE Trans. Pattern Analysis and Machine Intell., 42(2):328–342, 2018. 5, 6

    work page 2018

Show all 68 references
  1. [9]

    Pose determination from line-to-plane corre- spondences: existence condition and closed-form solutions

    Homer H Chen. Pose determination from line-to-plane corre- spondences: existence condition and closed-form solutions. IEEE Trans. Pattern Analysis and Machine Intell. , 13(06): 530–541, 1991. 2, 4

    1991

  2. [10]

    A plane-based approach for indoor point clouds reg- istration

    Ketty Favre, Muriel Pressigout, Eric Marchand, and Luce Morin. A plane-based approach for indoor point clouds reg- istration. In International Conference on Pattern Recogni- tion, pages 7072–7079, 2021. 6

    2021

  3. [11]

    Efficient and accurate registration of point clouds with plane to plane cor- respondences

    Wolfgang Forstner and Kourosh Khoshelham. Efficient and accurate registration of point clouds with plane to plane cor- respondences. In Proceedings of the IEEE international con- ference on computer vision workshops , pages 2165–2173,

  4. [12]

    The canonical cor- relations of matrix pairs and their numerical computation

    Gene H Golub and Hongyuan Zha. The canonical cor- relations of matrix pairs and their numerical computation . Springer, 1995. 3

    1995

  5. [13]

    PL-SLAM: A stereo SLAM system through the combination of points and line segments

    Ruben Gomez-Ojeda, Francisco-Angel Moreno, David Zuniga-No¨el, Davide Scaramuzza, and Javier Gonzalez- Jimenez. PL-SLAM: A stereo SLAM system through the combination of points and line segments. IEEE Trans. Robot., 35(3):734–746, 2019. 1

    2019

  6. [14]

    Multiple View Ge- ometry in Computer Vision

    Richard Hartley and Andrew Zisserman. Multiple View Ge- ometry in Computer Vision . Cambridge University Press,

  7. [15]

    Relations between two sets of variates

    Harold Hotelling. Relations between two sets of variates. In Breakthroughs in statistics: methodology and distribution , pages 162–190. 1992. 3

    1992

  8. [16]

    Learning to parse wireframes in images of man-made environments

    Kun Huang, Yifan Wang, Zihan Zhou, Tianjiao Ding, Shenghua Gao, and Yi Ma. Learning to parse wireframes in images of man-made environments. In Proc. IEEE Conf. on Comput. Vision and Pattern Recog., pages 626–635, 2018. 1

    2018

  9. [17]

    Projection metric learning on Grassmann manifold with application to video based face recognition

    Zhiwu Huang, Ruiping Wang, Shiguang Shan, and Xilin Chen. Projection metric learning on Grassmann manifold with application to video based face recognition. In Proc. IEEE Conf. on Comput. Vision and Pattern Recog. , pages 140–149, 2015. 2

    2015

  10. [18]

    Closed-form solutions for estimating a rigid motion from plane correspondences extracted from point clouds

    Kourosh Khoshelham. Closed-form solutions for estimating a rigid motion from plane correspondences extracted from point clouds. ISPRS Journal of Photogrammetry and Remote Sensing, 114:78–91, 2016. 6

    2016

  11. [19]

    Scalable dense non-rigid structure-from-motion: A Grassmannian perspective

    Suryansh Kumar, Anoop Cherian, Yuchao Dai, and Hong- dong Li. Scalable dense non-rigid structure-from-motion: A Grassmannian perspective. In Proc. IEEE Conf. on Comput. Vision and Pattern Recog., pages 254–263, 2018. 2

    2018

  12. [20]

    The grass- mannian of affine subspaces

    Lek-Heng Lim, Ken Sze-Wai Wong, and Ke Ye. The grass- mannian of affine subspaces. Foundations of Computational Mathematics, 21:537–574, 2021. 3

    2021

  13. [21]

    Pluck- ernet: Learn to register 3D line reconstructions

    Liu Liu, Hongdong Li, Haodong Yao, and Ruyi Zha. Pluck- ernet: Learn to register 3D line reconstructions. In Proc. IEEE Conf. on Comput. Vision and Pattern Recog. , pages 1842–1852, 2021. 1, 2, 4, 6, 7

    2021

  14. [22]

    3D line mapping revisited

    Shaohui Liu, Yifan Yu, R ´emi Pautrat, Marc Pollefeys, and Viktor Larsson. 3D line mapping revisited. In Proc. IEEE Conf. on Comput. Vision and Pattern Recog., pages 21445– 21455, 2023. 1

    2023

  15. [23]

    Globally op- timal camera orientation estimation from line correspon- dences by BnB algorithm

    Yinlong Liu, Guang Chen, and Alois Knoll. Globally op- timal camera orientation estimation from line correspon- dences by BnB algorithm. IEEE Trans. Robot. and Automat., 6(1):215–222, 2020. 4, 7, 8, 13

    2020

  16. [24]

    Global data association for SLAM with 3D Grassmannian manifold objects

    Parker C Lusk and Jonathan P How. Global data association for SLAM with 3D Grassmannian manifold objects. InProc. IEEE/RSJ Intl. Conf. on Intell. Robots and Sys., pages 4463– 4470, 2022. 1, 2, 8

    2022

  17. [25]

    Complete closed-form and accurate solution to pose estimation from 3D correspondences

    Ezio Malis. Complete closed-form and accurate solution to pose estimation from 3D correspondences. IEEE Robot. and Automat. Lett., 8(3):1786–1793, 2023. 2

    2023

  18. [26]

    A novel closed-form approach for enhancing ef- ficiency in pose estimation from 3D correspondences

    Ezio Malis. A novel closed-form approach for enhancing ef- ficiency in pose estimation from 3D correspondences. IEEE Robot. and Automat. Lett., 2024. 2

    2024

  19. [27]

    Globally op- timal pose estimation from line correspondences

    Faraz M Mirzaei and Stergios I Roumeliotis. Globally op- timal pose estimation from line correspondences. In Proc. IEEE Intl. Conf. on Robot. and Automat., pages 5581–5588,

  20. [28]

    ORB-SLAM: a versatile and accurate monocular SLAM system

    Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. ORB-SLAM: a versatile and accurate monocular SLAM system. IEEE Trans. Robot., 31(5):1147–1163, 2015. 1

    2015

  21. [29]

    Solving quadratically con- strained geometrical problems using Lagrangian duality

    Carl Olsson and Anders Eriksson. Solving quadratically con- strained geometrical problems using Lagrangian duality. In Proc. Intl. Conf. Pattern Recog., pages 1–5, 2008. 1, 2, 6, 7, 14 9

    2008

  22. [30]

    Branch- and-bound methods for Euclidean registration problems

    Carl Olsson, Fredrik Kahl, and Magnus Oskarsson. Branch- and-bound methods for Euclidean registration problems. IEEE Trans. Pattern Analysis and Machine Intell. , 31(5): 783–794, 2008. 1, 2, 5, 6

    2008

  23. [31]

    Colored point cloud registration revisited

    Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Colored point cloud registration revisited. In Proc. IEEE Intl. Conf. on Comput. Vision, pages 143–152, 2017. 6, 7, 14

    2017

  24. [32]

    Gluestick: Robust image matching by stick- ing points and lines together

    R ´emi Pautrat, Iago Su ´arez, Yifan Yu, Marc Pollefeys, and Viktor Larsson. Gluestick: Robust image matching by stick- ing points and lines together. In Proc. IEEE Intl. Conf. on Comput. Vision, pages 9706–9716, 2023. 7, 14

    2023

  25. [33]

    A review of point cloud registration algorithms for mobile robotics

    Franc ¸ois Pomerleau, Francis Colas, Roland Siegwart, et al. A review of point cloud registration algorithms for mobile robotics. Foundations and Trends® in Robotics, 4(1):1–104,

  26. [34]

    Structure- from-motion revisited

    Johannes L Schonberger and Jan-Michael Frahm. Structure- from-motion revisited. In Proc. IEEE Conf. on Comput. Vi- sion and Pattern Recog., pages 4104–4113, 2016. 1

    2016

  27. [35]

    Gen- eralized ICP

    Aleksandr Segal, Dirk Haehnel, and Sebastian Thrun. Gen- eralized ICP. In Proc. Robot.: Science & Sys. Conf. , page 435, 2009. 1

    2009

  28. [36]

    Object tracking via non-Euclidean ge- ometry: A Grassmann approach

    Sareh Shirazi, Mehrtash T Harandi, Brian C Lovell, and Conrad Sanderson. Object tracking via non-Euclidean ge- ometry: A Grassmann approach. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 901–908, 2014. 2

    2014

  29. [37]

    The replica dataset: A digital replica of indoor spaces

    Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797,

    1906 arXiv

  30. [38]

    A benchmark for the eval- uation of rgb-d slam systems

    J ¨urgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the eval- uation of rgb-d slam systems. In Proc. IEEE/RSJ Intl. Conf. on Intell. Robots and Sys., pages 573–580. IEEE, 2012. 7

    2012

  31. [39]

    Pose estimation from line correspondences: A complete analysis and a series of solutions

    Chi Xu, Lilian Zhang, Li Cheng, and Reinhard Koch. Pose estimation from line correspondences: A complete analysis and a series of solutions. IEEE Trans. Pattern Analysis and Machine Intell., 39(6):1209–1222, 2016. 7, 8, 13

    2016

  32. [40]

    Go-ICP: A globally optimal solution to 3D ICP point- set registration

    Jiaolong Yang, Hongdong Li, Dylan Campbell, and Yunde Jia. Go-ICP: A globally optimal solution to 3D ICP point- set registration. IEEE Trans. Pattern Analysis and Machine Intell., 38(11):2241–2254, 2015. 6, 8, 4

    2015

  33. [41]

    Grassmann manifold flows for stable shape generation

    Ryoma Yataka, Kazuki Hirashima, and Masashi Shiraishi. Grassmann manifold flows for stable shape generation. Ad- vances in Neural Information Processing Sys. Conf. , 36: 72377–72411, 2023. 2

    2023

  34. [42]

    Schubert varieties and distances between subspaces of different dimensions.SIAM Journal on Matrix Analysis and Applications, 37(3):1176–1197, 2016

    Ke Ye and Lek-Heng Lim. Schubert varieties and distances between subspaces of different dimensions.SIAM Journal on Matrix Analysis and Applications, 37(3):1176–1197, 2016. 3

    2016

  35. [43]

    Slim: Scalable and lightweight lidar mapping in urban environments

    Zehuan Yu, Zhijian Qiao, Wenyi Liu, Huan Yin, and Shaojie Shen. Slim: Scalable and lightweight lidar mapping in urban environments. arXiv preprint arXiv:2409.08681, 2024. 1, 8

    2024 arXiv

  36. [44]

    Grassmann graph embedding

    Bingxin Zhou, Xuebin Zheng, Yu Guang Wang, Ming Li, and Junbin Gao. Grassmann graph embedding. In ICLR workshop on geometrical and topological representation learning, 2021. 2

    2021

  37. [45]

    A com- plete, accurate and efficient solution for the perspective-n- line problem

    Lipu Zhou, Daniel Koppel, and Michael Kaess. A com- plete, accurate and efficient solution for the perspective-n- line problem. IEEE Trans. Robot. and Automat. , 6(2):699– 706, 2020. 7, 8, 13

    2020

  38. [46]

    Nice-slam: Neural implicit scalable encoding for SLAM

    Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hu- jun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Polle- feys. Nice-slam: Neural implicit scalable encoding for SLAM. In Proc. IEEE Conf. on Comput. Vision and Pattern Recog., pages 12786–12796, 2022. 6 10 Registration bey...

    2022

  39. [47]

    Proof of Theorem 1 Proof

    Proof of Theorems 7.1. Proof of Theorem 1 Proof. We need to show that f satisfies the following two properties of the group action. 1.∀X ∈ Gr(k, n), I · X = X (Identity ) 2.∀(X ∈ Gr(k, n), T1, T2 ∈ SE (n)), (T1T2) · X = T1 · (T2 · X) (Compatibility )

  40. [48]

    For the identity matrix I, it is straightforward that I · X = X, as the rotation matrix is the identity, and the translation is the zero vector: I · (A + b) = (I · A) + (Ib + I(I − AA⊤)I⊤0) (16) = A + b

  41. [49]

    Given two elements of SE (n), T1 = (R1, t1) and T2 = (R2, t2), T1 · (T2 · X) is derived by following process: T2 · X = (R2 · A) + R2b + R2(I − AA⊤)R⊤ 2 t2 (17) T1 · (T2 · X) = R1 · (R2 · A) + R1(R2b + R2(I − AA⊤)R⊤ 2 t2) + R1(I − R2A(R2A)⊤)R⊤ 1 t1 (18) = R1 · (R2 · A) + R1R2(b...

  42. [50]

    Grassmann Distance Inducing the geodesic distance on the Grassmannian requires a principal vector and angle defined as follows [42]: Definition 4 (Principal Vector)

    Derivation Details 8.1. Grassmann Distance Inducing the geodesic distance on the Grassmannian requires a principal vector and angle defined as follows [42]: Definition 4 (Principal Vector). Let A ∈ Gr(k, n), B ∈ Gr(l, n), and k ⩽ l < nbe positive integers. Then ith principal v...

  43. [51]

    for an arbitrary rotation R within the cube Cr using the triangle inequality of the Grassmann distance: dGr(Rdi 1, di

  44. [52]

    (44) Then from [40], an upper bound for dGr(R0di 1, Rdi

    − dGr(R0di 1, Rdi 1). (44) Then from [40], an upper bound for dGr(R0di 1, Rdi

  45. [53]

    is written as: dGr(R0di 1, Rdi

  46. [54]

    (45) From Eq

    ≤ min( π 2 , √ 3σr). (45) From Eq. (44) and Eq. (45), an upper bound for the objective function of Eq. (43) is derived as: max R∈Cr NX i=1 1 ϵ − dGr(Rdi 1, di 2)2 ≤ NX i=1 1 ϵ − max 0, dGr(R0di 1, di

  47. [55]

    (47) Additionally, a lower bound for the objective function in Eq

    − min( π 2 , √ 3σr) 2 (46) := ¯νr. (47) Additionally, a lower bound for the objective function in Eq. (43) is readily derived as: max R∈Cr NX i=1 1 ϵ − dGr(Rdi 1, di 2)2 ≥ NX i=1 1 ϵ − dGr(R0di 1, di 2)2 (48) := νr (49) 8.3.2. Line-to-plane case An objective function for this ...

  48. [56]

    Then, from the triangle inequality of Euclidean distance, lower bound of Pz(T · li

    ˜bi 2 − ˜bi 2 2 2 , (58) where T = ( R∗, t) and Pz(T · li 1)bi 2 = (˜bi 2)⊤R∗di 1 R∗di 1 + (˜bi 2)⊤ ˜b ′i 1 (R∗, t) ˜b ′i 1 (R∗, t). Then, from the triangle inequality of Euclidean distance, lower bound of Pz(T · li

  49. [57]

    ˜bi 2 − ˜bi 2 2 is: Pz(T · li

  50. [58]

    ˜bi 2 − ˜bi 2 2 ≥ Pz(T0 · li

  51. [59]

    ˜bi 2 − ˜bi 2 2 − Pz(T0 · li

  52. [60]

    From its definition, Pz(T0 · li

    ˜bi 2 2 , (59) where T0 = (R∗, t0). From its definition, Pz(T0 · li

  53. [61]

    ˜bi 2 2 can be rewritten as: Pz(T0 · li

  54. [62]

    (60) Recall that from Theorem 3: b ′i 1 (R∗, t) = R∗bi 1 + R∗(I − dd⊤)R∗⊤t

    ˜bi 2 2 = (˜bi 2)⊤ ˜b ′i 1 (R∗, t0) ˜b ′i 1 (R∗, t0) − (˜bi 2)⊤ ˜b ′i 1 (R∗, t) ˜b ′i 1 (R∗, t) 2 . (60) Recall that from Theorem 3: b ′i 1 (R∗, t) = R∗bi 1 + R∗(I − dd⊤)R∗⊤t. (61) Since t ∈ Ct, a set of vectors b ′i 1 (R∗, t) forms a line segment within R3, where its two end-...

  55. [63]

    ˜bi 2 − ˜bi 2 2 2 ≥ NX i=1 max(0, Pz(T0 · li

  56. [64]

    ˜bi 2 − ˜bi 2 2 − ψt) 2 (63) = et (64) Also, an upper bound is: min t∈Ct NX i=1 Pz(T · li

  57. [65]

    ˜bi 2 − ˜bi 2 2 2 ≤ NX i=1 Pz(T0 · li

  58. [66]

    ˜bi 2 − ˜bi 2 2 2 (65) = ¯et (66) The process for obtaining the bounds is exactly the same for the case of the line-to-plane and plane-to-plane cases. 6 8.5. Algorithms This section introduces the entire pipeline for solving the line-to-line registration problem with our BnB s...

  59. [67]

    Measurement Variation of Point-based Registration In this section, we further analyze how the optimal rotation of point-based cost functions varies with changes in point location

    Analysis on Point-based and Parameter-based Methods 9.1. Measurement Variation of Point-based Registration In this section, we further analyze how the optimal rotation of point-based cost functions varies with changes in point location. For simplicity, our analysis focuses on ...

  60. [68]

    Time Complexity Analysis In this section, we provide a computational time of experiments in Sec

    Experiments Details 10.1. Time Complexity Analysis In this section, we provide a computational time of experiments in Sec. 5. All the reported times represent the average time required to process a single set. For example, in the object registration experiment, the time for a ...

Pith tools

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