Pith. sign in

REVIEW 4 major objections 6 minor 50 references

Fully-Geometric Cross-Attention for Point Cloud Registration

T0 review · 4 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read Enriching cross-attention with transformation-invariant distance and angle information makes point cloud registration more accurate, especially when overlap is low.

desk verdict A genuinely new geometric cross-attention with Gromov-Wasserstein embeddings that delivers real gains on low-overlap registration, but the signed-eigenvector stability assumption needs an explicit check. read the letter →

arxiv 2502.08285 v2 pith:N2JWDL6Y submitted 2025-02-12 cs.CV

classification cs.CV
keywords pointcloudregistrationcross-attentionGromov-Wassersteindistancetransformationinvariancesuper-pointmatchingcoarse-to-finelow-overlapTransformer
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 cross-attention between two point clouds should use coordinate information as well as learned features, and that this can be done while preserving rotation and translation invariance. The proposed method, FLAT, augments the attention score with a pair-wise distance embedding computed from the Gromov-Wasserstein distance between local geometric structures and a triplet-wise angle embedding computed across the two clouds, so that points can attend to each other under arbitrary rigid transformations. At the super-point level this produces more distinctive features and more inlier correspondences; a distance-weighted self-attention at the point level then sharpens fine matching. If the claim holds, low-overlap and cross-source registration become more reliable, which matters for robotics, autonomous driving, and 3D reconstruction.

What carries the argument

The machinery is the fully-geometric cross-attention block, which the paper calls FLAT. It replaces the plain attention score by the sum of feature similarity and a geometric structure embedding made of two transformation-invariant pieces: a pair-wise cross-distance embedding derived from the Gromov-Wasserstein distance between the local metric structures of two super-points, and a triplet-wise cross-angle embedding computed between feature vectors that concatenate eigenvalue information. The signed eigenvector of the smallest eigenvalue of each local covariance matrix supplies the orientation reference that makes the angle term well-defined across clouds; the geometric embedding is what lets the network emphasize overlapping regions and learn corresponding structures.

What would settle it

Take the 3DLoMatch test pairs, group super-points by the gap between the two smallest eigenvalues of their local covariance, and compare FLAT's inlier ratio against a version with the eigenvector term removed or sign-flipped. If the advantage over vanilla cross-attention disappears exactly on low-gap, planar super-points, the reported gains are driven by the eigenvector orientation assumption rather than by the distance embedding; if the advantage persists, the orientation reference is not actually load-bearing.

Watch

Extended reading notes

Core claim

The central claim is that feeding geometric structure into cross-attention, not just feature similarity, is what makes Transformer-based registration accurate on hard pairs, and that the necessary geometry can be made invariant to the independent reference frames of the two clouds. FLAT implements this with a cross-attention score in which a learned sinusoidal embedding of a Gromov-Wasserstein distance between local covariance structures, plus a triplet-wise angle defined from signed eigenvectors and eigenvalue-based feature vectors, is added to the standard query-key product. The paper reports that this geometry-enhanced attention, combined with a distance-weighted local self-attention and an optimal-transport coarse-to-fine matching pipeline, raises inlier ratios and registration recall relative to strong baselines, with the largest gains on the low-overlap 3DLoMatch benchmark and on the cross-source 3DCSR benchmark.

Load-bearing premise

The method relies on a consistently oriented local direction, the signed eigenvector of each neighborhood's covariance, staying stable in every region; on flat or repetitive surfaces that direction can flip essentially at random, and the attention weights built on it inherit the resulting noise.

Editorial extensions

If this is right

  • On 3DLoMatch, the low-overlap indoor benchmark, FLAT reports registration recall of 78.7% versus 75.0% for GeoTransformer at 5000 samples, the largest gap among the compared methods.
  • On 3DMatch, the full geometry-augmented cross-attention raises inlier ratio from 73.4% (vanilla cross-attention) to 84.2%, meaning fewer false correspondences reach the transformation estimator.
  • A model trained only on 3DMatch reaches 62.9% registration recall on the cross-source 3DCSR benchmark, beating 60.2% for GeoTransformer, which supports the claim that the geometric attention transfers across sensors.
  • On KITTI, FLAT matches the best registration recall (99.8%) while achieving the lowest relative rotation error (0.24 degrees), indicating the gains carry over to outdoor LiDAR data.
  • Ablations show that both the pair-wise distance embedding and the triplet-wise angle embedding contribute, with the combined version outperforming either alone.

Reading between the lines

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

  • The signed eigenvector of the smallest local eigenvalue is the fragile point: on planar or symmetric neighborhoods the two smallest eigenvalues nearly coincide, so the eigenvector direction and sign become noise-dominated, and that noise propagates into both the Gromov-Wasserstein cost and the angle embedding. A testable variant would replace or smooth the eigenvector when the eigenvalue gap is sm
  • The paper's own Limitations section says the pair-wise distance mapping is relatively expensive and that density differences between source and target clouds remain a challenge; both point to the next practical bottleneck, namely making the geometry cost cheaper and robust to different sensor resolutions.
  • Because the attention maps visibly concentrate on overlapping regions, the learned cross-attention could be reused as a free overlap prior for other tasks such as overlap prediction or partial-to-partial matching.
  • A simple ablation of the term $\alpha\|v_{\bar p_s}-v_{\bar p_k}\|_2^2$ in Eq. (4) would isolate how much of the reported gain comes from orientation consistency between eigenvector directions rather than from the distance structure alone.
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 / 6 minor

Summary. The paper proposes FLAT, a coarse-to-fine deep network for rigid point cloud registration. Its main novelty is a 'fully-geometric' cross-attention mechanism that augments standard feature-based cross-attention with transformation-invariant geometric embeddings: a pairwise distance embedding computed from a Gromov-Wasserstein optimal transport plan between superpoints, where the intra-cloud costs include a signed-eigenvector term, and a triplet-wise angular embedding computed from eigenvalue features. A distance-weighted local self-attention is also introduced for fine-level matching. The method is evaluated on 3DMatch, 3DLoMatch, KITTI, and 3DCSR, reporting state-of-the-art or competitive results, with an ablation study attributing large inlier-ratio gains to the geometric cross-attention.

Significance. If the technical gaps are resolved, the paper would make a solid contribution: it demonstrates a plausible way to inject coordinate information into cross-attention while preserving rigid invariance, which is a recognized limitation of prior feature-only cross-attention. The reported gains on low-overlap indoor benchmarks are meaningful, and the promise of code will facilitate reproducibility. However, the robustness of the signed-eigenvector encoding and the interpretation of the ablation results are not yet established, and one reported KITTI result is contradicted by the paper's own table.

major comments (4)
  1. [Eq. (4) and Section 3.3] The signed eigenvector v associated with the smallest eigenvalue of the local covariance matrix in Eq. (3) is used in the cost matrices C^p and C^q of the Gromov-Wasserstein distance in Eq. (4). The sign rule ('positive from the center of the point cloud towards the current point') is not formally defined and can be inconsistent on planar or nearly symmetric neighborhoods where λ1 ≈ λ2; in such cases the eigenvector direction is noise-dominated and the sign can flip between neighboring superpoints. Because the pairwise distance embedding derived from the transport plan S contributes the largest ablation gain (Table 4: IR from 73.4 to 81.4 on 3DMatch and 43.8 to 54.4 on 3DLoMatch), this is a load-bearing assumption. The paper should provide statistics on eigenvalue gaps for the datasets, a sensitivity analysis for the learned parameter α, and an analysis of sign-flip rates, or replace v with a sign-invariant descriptor such as v v^T.
  2. [Table 4] The ablation table is ambiguous: the rows 'cross-attention w/DAE' and 'self-attention w/PDE' report identical values (92.2/98.1/84.2 and 78.1/89.6/57.3) on all six metrics. This suggests the two rows are not independent ablations, yet the text describes them as evaluating different components. Without a clear statement of the baseline architecture in each section and a full 2x2 ablation (with/without geometric cross-attention and with/without distance-weighted self-attention), the central claim that both mechanisms contribute to the improvement is not supported.
  3. [Section 4.2 and Table 2] The text states that FLAT achieves 'the lowest average RTE and RRE' on KITTI, but Table 2 shows Predator with RTE 6.8 cm versus FLAT 6.9 cm. The claim about RTE is thus false as written, and the paper should correct it or explain why the comparison is not directly applicable.
  4. [Eq. (4) and Pair-wise Cross Distance Embedding] The paper defines S as the Gromov-Wasserstein coupling and then uses r^D_ij = sinu((1 - s_ij)/σ_ρ). A raw optimal transport plan from Sinkhorn is a joint distribution with entries typically of order 1/(N M) (or otherwise dependent on the mass normalization); without specifying how S is scaled or normalized before forming 1 - s_ij, the sinusoidal embedding may be dominated by values near 1 for all pairs, which would defeat the purpose of the distance embedding. The authors should specify the normalization and show that the resulting embedding carries discriminative information.
minor comments (6)
  1. [Eq. (4)] The second constraint line writes C^q_{sk} but the right-hand side uses \bar{q}_t and \bar{q}_l; this should be C^q_{tl}.
  2. [Section 3.3.1, Coarse Matching] The first norm in \bar{C}_{ij} uses \hat{f}_{\bar{p}_i} and \hat{f}_{\bar{p}_j}; the second argument should be \hat{f}_{\bar{q}_j}.
  3. [Section 3.3, Local Geometric Self-attention] The definition of d^i_{kl} contains a set-notation error; it should read d^i_{kl} = \|\bar{p}_{ik} - \bar{p}_{il}\|_2^2.
  4. [References] Reference [16] is cited for the 3DCSR dataset, but [16] is a survey paper; the original 3DCSR dataset paper should be cited instead.
  5. [Section 4.4] The sentence reporting '2.8% (81.4% vs. 84.2%) IR' is confusing because the vanilla-to-DAE IR gain is 73.4 to 84.2 (10.8 points); the authors should clarify that the quoted numbers are the incremental gains from adding TAE.
  6. [Figure 6 caption] The description of the red rectangles and the two configurations is hard to follow; please clarify what is shown in each row.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: FLAT's central claim is an externally evaluated empirical improvement, and its geometric embeddings are computed from inputs (coordinates and local covariance eigendecompositions) rather than from the prediction targets.

full rationale

The paper's load-bearing claim is that a fully-geometric cross-attention, using transformation-invariant distance and angle embeddings, improves inlier ratio and registration recall on 3DMatch, 3DLoMatch, KITTI, and 3DCSR (Abstract, Section 1, Table 1-3). This claim is tested against held-out scenes with ground-truth poses, and the method is trained with supervised losses (Eq. 13 and fine matching loss) that use ground-truth overlap only as supervision, not as inputs to the geometric embedding. The cross-geometric distance in Eq. (4) is computed from coordinates and local covariance eigenvector differences, and the angle embedding in Eq. (5) is computed from eigenvalues and a global eigenvalue center; neither reduces to the target registration error or to a fitted constant that is later reported as a prediction. The ablation in Table 4 attributes gains to the proposed PDE/TAE/DAE modules, which is a legitimate internal comparison rather than a circular derivation. The reader's concern about signed-eigenvector stability (λ1 ≈ λ2 and sign flips on near-symmetric neighborhoods) is a real robustness/correctness risk, but it does not make the derivation circular: it means a geometric ingredient may be unstable for some inputs, not that the method's output is its own input by construction. The paper cites prior work by its own authors in the related-work and reference sections, but these citations are not load-bearing: the core novelty is not justified by a self-citation chain, and no uniqueness theorem is imported from the authors' own prior work. The method is self-contained against external benchmarks, so the honest finding is no significant circularity.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

The central claim rests on four main imported or assumed pieces: rigid invariance of covariance eigenvalues, stability of a signed eigenvector, the use of a Gromov-Wasserstein coupling as an attention similarity, and the geometric meaning of an angle defined in an eigenvalue feature space. The signed-eigenvector stability is the most fragile because it is load-bearing for both the distance and angle embeddings and is not analyzed.

free parameters (8)
  • alpha (α) in Eq. 4 = learned
    Weights the eigenvector difference term inside the cross-cloud Gromov-Wasserstein cost matrices.
  • sigma_rho (σ_ρ) = learned
    Scales the pair-wise cross distance embedding r^D_ij in Eq. 6.
  • sigma_theta (σ_θ) = learned
    Scales the triplet-wise cross angle embedding r^A_ijx.
  • eta (η) = 0.1
    Balances feature distance and structure distance in the coarse matching cost matrix in Section 3.3.1.
  • tau_c = 0.15
    Confidence threshold for keeping super-point correspondences; set per dataset.
  • radii r, r_o, r_p, r_n = 3.75 cm, 45 cm, 21 cm, 75 cm
    Neighborhood, overlap, and negative radii for indoor and KITTI; inherited from prior protocols.
  • omega (ω) = learned
    Scale factor in the circle loss for fine matching.
  • neighbor count k = 10
    Number of neighbors sampled in the ball for covariance estimation in Eq. 3.
assumptions (4)
  • standard math Eigenvalues of a local covariance matrix are invariant to rigid transformations.
    Used after Eq. 3 to justify that the eigenvalue tuple λ and eigenvector v encode geometry independent of the reference frame.
  • domain assumption The signed eigenvector v has a well-defined and stable orientation for every superpoint.
    The paper signs v to point from the cloud center to the current point, but on symmetric or planar neighborhoods the smallest eigenvector is ambiguous; no stability analysis is provided.
  • domain assumption The Gromov-Wasserstein coupling S between superpoint clouds is a valid point-to-point similarity for attention.
    Equations 4-6 convert the coupling into distance embeddings r^D_ij without justification that S values are comparable across different scenes.
  • ad hoc to paper Angles computed in the concatenated eigenvalue and global-center feature space are meaningful cross-cloud geometric signals.
    Eq. 5 defines an angle in a feature space that mixes local eigenvalues with a global center; no derivation shows this is a metric or geometrically interpretable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fully-Geometric Cross-Attention for Point Cloud Registration." pith.science (2026). https://pith.science/paper/N2JWDL6Y

@misc{pith2026250208285,
  author       = {Pith},
  title        = {Pith review of: Fully-Geometric Cross-Attention for Point Cloud Registration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N2JWDL6Y}},
  note         = {Machine review of arXiv:2502.08285}
}
read the original abstract

Point cloud registration approaches often fail when the overlap between point clouds is low due to noisy point correspondences. This work introduces a novel cross-attention mechanism tailored for Transformer-based architectures that tackles this problem, by fusing information from coordinates and features at the super-point level between point clouds. This formulation has remained unexplored primarily because it must guarantee rotation and translation invariance since point clouds reside in different and independent reference frames. We integrate the Gromov-Wasserstein distance into the cross-attention formulation to jointly compute distances between points across different point clouds and account for their geometric structure. By doing so, points from two distinct point clouds can attend to each other under arbitrary rigid transformations. At the point level, we also devise a self-attention mechanism that aggregates the local geometric structure information into point features for fine matching. Our formulation boosts the number of inlier correspondences, thereby yielding more precise registration results compared to state-of-the-art approaches. We have conducted an extensive evaluation on 3DMatch, 3DLoMatch, KITTI, and 3DCSR datasets.

Figures

Figures reproduced from arXiv: 2502.08285 by the authors.

Figure 1
Figure 1. The encoder downsizes the input point clouds and generates super-points with associated features. The global geometric Trans [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Angle and distance provide geometry information for [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Examples of qualitative registration results on the [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Examples of qualitative registration results on the 3DLo [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 6
Figure 6. Figure 6: Visualization of attention weights for the point indicated [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 47 canonical work pages

  1. [1]

    Spinnet: Learning a general surface descrip- tor for 3d point cloud registration

    Sheng Ao, Qingyong Hu, Bo Yang, Andrew Markham, and Yulan Guo. Spinnet: Learning a general surface descrip- tor for 3d point cloud registration. InCVPR, pages 11753– 11762, 2021. 2, 6, 7

  2. [2]

    Pointnetlk: Robust & efficient point cloud registration using pointnet

    Yasuhiro Aoki, , and et al. Pointnetlk: Robust & efficient point cloud registration using pointnet. InCVPR, pages 7163–7172, 2019. 1

  3. [3]

    D3feat: Joint learning of dense de- tection and description of 3d local features

    Xuyang Bai and et al. D3feat: Joint learning of dense de- tection and description of 3d local features. InCVPR, pages 6359–6367, 2020. 2, 6, 7

  4. [4]

    Pointdsc: Robust point cloud registra- tion using deep spatial consistency

    Xuyang Bai and et al. Pointdsc: Robust point cloud registra- tion using deep spatial consistency. InCVPR, pages 15859– 15869, 2021. 1

  5. [5]

    Point cloud registration: a mini-review of current state, challenging is- sues and future directions.AIMS Geosciences, 9(1):68–85,

    Nathan Brightman, Lei Fan, and Yang Zhao. Point cloud registration: a mini-review of current state, challenging is- sues and future directions.AIMS Geosciences, 9(1):68–85,

  6. [6]

    3d shape knowledge graph for cross-domain 3d shape retrieval.CAAI Transactions on Intelligence Tech- nology, 9(5):1199–1216, 2024

    Rihao Chang, Yongtao Ma, Tong Hao, Weijie Wang, and Weizhi Nie. 3d shape knowledge graph for cross-domain 3d shape retrieval.CAAI Transactions on Intelligence Tech- nology, 9(5):1199–1216, 2024. 1

  7. [7]

    Deep global registration

    Christopher Choy and et al. Deep global registration. In CVPR, pages 2514–2523, 2020. 1, 2, 7

  8. [8]

    Fully convolutional geometric features

    Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. InICCV, pages 8958– 8966, 2019. 1, 2, 6, 7

Show all 50 references
  1. [9]

    Corsetti, D

    J. Corsetti, D. Boscaini, and F. Poiesi. Revisiting fully con- volutional geometric features for object 6d pose estimation. InICCVW, 2023. 1

  2. [10]

    Sinkhorn distances: Lightspeed computation of optimal transport.NeurIPS, 26:2292–2300, 2013

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport.NeurIPS, 26:2292–2300, 2013. 5

  3. [11]

    Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography.COMMUN ACM, 24(6):381–395, 1981

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

  4. [12]

    Robust point cloud registration frame- work based on deep graph matching

    Kexue Fu and et al. Robust point cloud registration frame- work based on deep graph matching. InCVPR, pages 8893– 8902, 2021. 2, 5

  5. [13]

    Are we ready for autonomous driving? the kitti vision benchmark suite

    Andreas Geiger and et al. Are we ready for autonomous driving? the kitti vision benchmark suite. InCVPR, pages 3354–3361. IEEE, 2012. 2, 6

  6. [14]

    Predator: Registration of 3d point clouds with low overlap

    Shengyu Huang and et al. Predator: Registration of 3d point clouds with low overlap. InCVPR, pages 4267–4276, 2021. 1, 2, 3, 6, 7

  7. [15]

    Feature- metric registration: A fast semi-supervised approach for ro- bust point cloud registration without correspondences

    Xiaoshui Huang, Guofeng Mei, and Jian Zhang. Feature- metric registration: A fast semi-supervised approach for ro- bust point cloud registration without correspondences. In CVPR, pages 11366–11374, 2020. 1

  8. [16]

    A comprehensive survey on point cloud registration

    Xiaoshui Huang, Guofeng Mei, Jian Zhang, and Rana Ab- bas. A comprehensive survey on point cloud registration. arXiv preprint arXiv:2103.02690, 2021. 2, 6, 7

  9. [17]

    Unsupervised point cloud regis- tration by learning unified gaussian mixture models.RA-L, 7 (3):7028–7035, 2022

    Xiaoshui Huang, Sheng Li, Yifan Zuo, Yuming Fang, Jian Zhang, and Xiaowei Zhao. Unsupervised point cloud regis- tration by learning unified gaussian mixture models.RA-L, 7 (3):7028–7035, 2022. 1

  10. [18]

    Slam-driven robotic mapping and registration of 3d point clouds.Au- tomation in Construction, 89:38–48, 2018

    Pileun Kim, Jingdao Chen, and Yong K Cho. Slam-driven robotic mapping and registration of 3d point clouds.Au- tomation in Construction, 89:38–48, 2018. 1

  11. [19]

    Freeinsert: Disentangled text-guided object insertion in 3d gaussian scene without spatial priors.arXiv preprint arXiv:2505.01322, 2025

    Chenxi Li, Weijie Wang, Qiang Li, Bruno Lepri, Nicu Sebe, and Weizhi Nie. Freeinsert: Disentangled text-guided object insertion in 3d gaussian scene without spatial priors.arXiv preprint arXiv:2505.01322, 2025. 1

  12. [20]

    Iterative distance- aware similarity matrix convolution with mutual-supervised point elimination for efficient point cloud registration

    Jiahao Li, Changhao Zhang, and et al. Iterative distance- aware similarity matrix convolution with mutual-supervised point elimination for efficient point cloud registration. In ECCV, 2019. 2

  13. [21]

    Point cloud registration with self-supervised feature learning and beam search

    Guofeng Mei. Point cloud registration with self-supervised feature learning and beam search. InDICTA, pages 01–08,

  14. [22]

    Overlap-guided coarse-to-fine correspondence prediction for point cloud registration

    Guofeng Mei, Xiaoshui Huang, Jian Zhang, and Qiang Wu. Overlap-guided coarse-to-fine correspondence prediction for point cloud registration. InICME, pages 1–6. IEEE, 2022. 1

  15. [23]

    G. Mei, C. Saltori, F. Poiesi, J. Zhang, E. Ricci, N. Sebe, and Q. Wu. Data augmentation-free unsupervised learning for 3d point cloud understanding. InBMVC, 2022. 1

  16. [24]

    Overlap-guided gaussian mix- ture models for point cloud registration

    Guofeng Mei, Fabio Poiesi, Cristiano Saltori, Jian Zhang, Elisa Ricci, and Nicu Sebe. Overlap-guided gaussian mix- ture models for point cloud registration. InWACV, pages 4511–4520, 2023. 1

  17. [25]

    Unsu- pervised deep probabilistic approach for partial point cloud registration

    Guofeng Mei, Hao Tang, Xiaoshui Huang, Weijie Wang, Juan Liu, Jian Zhang, Luc Van Gool, and Qiang Wu. Unsu- pervised deep probabilistic approach for partial point cloud registration. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 136...

  18. [26]

    Hgan: Holistic generative adversarial networks for two- dimensional image-based three-dimensional object retrieval

    Weizhi Nie, Weijie Wang, Anan Liu, Jie Nie, and Yuting Su. Hgan: Holistic generative adversarial networks for two- dimensional image-based three-dimensional object retrieval. ACM Transactions on Multimedia Computing, Communica- tions, and Applications (TOMM), 15(4):1–24, 2019. 1

  19. [27]

    T2td: Text-3d generation model based on prior knowledge guidance.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    Weizhi Nie, Ruidong Chen, Weijie Wang, Bruno Lepri, and Nicu Sebe. T2td: Text-3d generation model based on prior knowledge guidance.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1

  20. [28]

    3dregnet: A deep neural network for 3d point registration

    G Dias Pais, Srikumar Ramalingam, Venu Madhav Govindu, Jacinto C Nascimento, Rama Chellappa, and Pedro Miraldo. 3dregnet: A deep neural network for 3d point registration. In CVPR, pages 7193–7203, 2020. 2

  21. [29]

    Computational optimal transport: With applications to data science.F oundations and Trends® in Machine Learning, 11(5-6):355–607, 2019

    Gabriel Peyr ´e, Marco Cuturi, et al. Computational optimal transport: With applications to data science.F oundations and Trends® in Machine Learning, 11(5-6):355–607, 2019. 3, 4

  22. [30]

    Poiesi and D

    F. Poiesi and D. Boscaini. Generalisable and distinc- tive 3D local deep descriptors for point cloud registration. arXiv:2105.10382, 2021. 1, 2

  23. [31]

    Geometric transformer for fast and ro- bust point cloud registration

    Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and ro- bust point cloud registration. InCVPR, pages 11143–11152,

  24. [32]

    Kpconv: Flexible and deformable convolution for point clouds

    Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatriz Marcotegui, Franc ¸ois Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. InProceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 3, 4

  25. [33]

    Attention is all you need.NeurIPS, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.NeurIPS, 30, 2017. 1, 2, 3, 4

  26. [34]

    You only hypothesize once: Point cloud registration with rotation-equivariant descriptors

    Haiping Wang, Yuan Liu, Zhen Dong, and Wenping Wang. You only hypothesize once: Point cloud registration with rotation-equivariant descriptors. InACM MM, pages 1630– 1641, 2022. 2, 6

  27. [35]

    Roreg: Pairwise point cloud registration with oriented descriptors and local rotations.TPAMI, 2023

    Haiping Wang, Yuan Liu, Qingyong Hu, Bing Wang, Jian- guo Chen, Zhen Dong, Yulan Guo, Wenping Wang, and Bisheng Yang. Roreg: Pairwise point cloud registration with oriented descriptors and local rotations.TPAMI, 2023. 1, 2

  28. [36]

    Zeroreg: Zero-shot point cloud registration with foundation models,

    Weijie Wang, Wenqi Ren, Guofeng Mei, Bin Ren, Xiaoshui Huang, Fabio Poiesi, Nicu Sebe, and Bruno Lepri. Zeroreg: Zero-shot point cloud registration with foundation models,

  29. [37]

    Uvmap-id: A controllable and personalized uv map generative model

    Weijie Wang, Jichao Zhang, Chang Liu, Xia Li, Xingqian Xu, Humphrey Shi, Nicu Sebe, and Bruno Lepri. Uvmap-id: A controllable and personalized uv map generative model. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 10725–10734, 2024. 1

  30. [38]

    Deep closest point: Learn- ing representations for point cloud registration

    Yue Wang and Justin M Solomon. Deep closest point: Learn- ing representations for point cloud registration. InICCV, pages 3523–3532, 2019. 2

  31. [39]

    Prnet: Self-supervised learning for partial-to-partial registration

    Yue Wang and Justin M Solomon. Prnet: Self-supervised learning for partial-to-partial registration. InNeurIPS, 2019. 3

  32. [40]

    A bayesian regularization network approach to thermal distortion control in 3d printing.Computational Me- chanics, pages 1–18, 2023

    Yuxi Xie, Boyuan Li, Chao Wang, Kun Zhou, CT Wu, and Shaofan Li. A bayesian regularization network approach to thermal distortion control in 3d printing.Computational Me- chanics, pages 1–18, 2023. 1

  33. [41]

    Omnet: Learning overlapping mask for partial- to-partial point cloud registration

    Hao Xu, Shuaicheng Liu, Guangfu Wang, Guanghui Liu, and Bing Zeng. Omnet: Learning overlapping mask for partial- to-partial point cloud registration. InICCV, pages 3132– 3141, 2021. 1

  34. [42]

    Glorn: Strong generalization fully convolutional network for low- overlap point cloud registration.T-GE, 60:1–14, 2022

    Jiabo Xu, Yukun Huang, Zeyun Wan, and Jingbo Wei. Glorn: Strong generalization fully convolutional network for low- overlap point cloud registration.T-GE, 60:1–14, 2022. 1, 6

  35. [43]

    Rpm-net: Robust point matching using learned features

    Zi Jian Yew and Gim Hee Lee. Rpm-net: Robust point matching using learned features. InCVPR, pages 11824– 11833, 2020. 2

  36. [44]

    Regtr: End-to-end point cloud correspondences with transformers

    Zi Jian Yew and Gim Hee Lee. Regtr: End-to-end point cloud correspondences with transformers. InCVPR, pages 6677–6686, 2022. 1, 3

  37. [45]

    Cofinet: Reliable coarse-to-fine correspon- dences for robust pointcloud registration.NeurIPS, 34, 2021

    Hao Yu and et al. Cofinet: Reliable coarse-to-fine correspon- dences for robust pointcloud registration.NeurIPS, 34, 2021. 1, 2, 4, 5, 6, 7

  38. [46]

    Rotation-invariant transformer for point cloud matching

    Hao Yu, Zheng Qin, Ji Hou, Mahdi Saleh, Dongsheng Li, Benjamin Busam, and Slobodan Ilic. Rotation-invariant transformer for point cloud matching. InCVPR, pages 5384– 5393, 2023. 6

  39. [47]

    3dmatch: Learning local geometric descriptors from rgb-d reconstruc- tions

    Andy Zeng, Shuran Song, Matthias Nießner, Matthew Fisher, Jianxiong Xiao, and Thomas Funkhouser. 3dmatch: Learning local geometric descriptors from rgb-d reconstruc- tions. InCVPR, pages 1802–1811, 2017. 2, 6

  40. [48]

    Patchformer: An efficient point transformer with patch at- tention

    Cheng Zhang, Haocheng Wan, Xinyi Shen, and Zizhao Wu. Patchformer: An efficient point transformer with patch at- tention. InCVPR, pages 11799–11808, 2022. 2

  41. [49]

    Deep learning based point cloud registration: an overview.VRIH, 2(3):222– 246, 2020

    Zhiyuan Zhang, Yuchao Dai, and Jiadai Sun. Deep learning based point cloud registration: an overview.VRIH, 2(3):222– 246, 2020. 1

  42. [50]

    Open3d: A modern library for 3d data processing.arXiv preprint arXiv:1801.09847, 2018

    Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3d: A modern library for 3d data processing.arXiv preprint arXiv:1801.09847, 2018. 5

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.