Pith. sign in

REVIEW 4 major objections 6 minor 61 references

DefTransNet: A Transformer-based Method for Non-Rigid Point Cloud Registration in the Simulation of Soft Tissue Deformation

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

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

Pith's one-line read The paper presents DefTransNet, a Transformer-based non-rigid point cloud registration network that it claims consistently beats state-of-the-art methods on four datasets, including real soft-tissue deformation data.

desk verdict A re-implementable combination of known components that deserves review, but the 'consistently outperforms SOTA' claim is contradicted by its own tables and the 4DMatch benchmark is modified in a way that weakens the partial-overlap test. read the letter →

arxiv 2502.06336 v1 pith:GG3FB3CJ submitted 2025-02-10 cs.CV cs.AI

Sara Monji-Azad , Marvin Kinz , Siddharth Kothari , Robin Khanna , Amrei Carla Mihan , David Maennel , Claudia Scherl , Juergen Hesser This is my paper · ORCID
classification cs.CVcs.AI
keywords non-rigidpointcloudregistrationTransformersofttissuedeformationdisplacementfieldfeatureambiguityLoopyBeliefPropagationEdgeConvChamferdistance
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

DefTransNet is a proposed end-to-end Transformer-based network for non-rigid point cloud registration, aimed at tracking deforming soft tissue during surgery. The paper argues that combining a learnable affine alignment, EdgeConv local/global geometric features, and a Transformer encoder-decoder resolves feature ambiguity and makes registration robust to large deformations, noise, outliers, and partial overlap. In experiments on SynBench, ModelNet, DeformedTissue, and 4DMatch, the authors report that DefTransNet achieves lower mean distance and Chamfer distance errors than several published baselines at nearly every condition tested. If these results hold, the method would offer a more accurate way to update the location and shape of anatomical structures during tumor resections.

What carries the argument

The load-bearing components are (1) a learnable transformation matrix, derived from a global feature vector via an MLP, that standardizes the two clouds before feature extraction; (2) EdgeConv layers that build a local graph and aggregate both global coordinates and local offsets; and (3) a Transformer encoder-decoder with multi-head self-attention, where the decoder output is added residually to the initial features to give source and target features $\Phi_X$ and $\Phi_Y$. A displacement network then finds $k$ nearest neighbors in the target for each source point, computes a feature-distance cost, adds a graph regularization term, and refines the costs with Loopy Belief Propagation before taking a softmax-weighted combination of candidate displacements. The paper contends that this pipeline, by capturing long-range dependencies and preserving local geometry, reduces feature ambiguity and yields robust non-rigid alignment.

What would settle it

Run DefTransNet on the original 4DMatch and 4DLoMatch targets without the equal-cardinality reconstruction; if its Chamfer distances no longer beat the baselines, the claimed low-overlap advantage is refuted.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a feature descriptor which first aligns input clouds with a learnable 3x3 transformation matrix, then pools local EdgeConv edge features into a global vector, and then enriches both source and target features with a Transformer encoder-decoder whose output is added back to the input features, produces displacement fields accurate enough to outperform current non-rigid registration networks. The network outputs a displacement vector field rather than an explicit correspondence set, using k-NN candidate search in the target, Loopy Belief Propagation message passing to enforce spatial consistency, and softmax weighting of candidate displacements. The authors claim this design “consistently outperforms the state-of-the-art methods across all datasets and deformation levels,” with the largest margins under heavy deformation, noise, outliers, and low overlap.

Load-bearing premise

The load-bearing premise is that the modified 4DMatch target, which is rebuilt to have the same number of points as the source, still measures the partial-overlap robustness the paper claims.

Editorial extensions

If this is right

  • If DefTransNet's accuracy holds, it could be used intraoperatively to track tissue shift during head-and-neck tumor resections, giving surgeons updated locations of critical structures.
  • The reported robustness to Gaussian noise and outliers suggests the network can operate on scans with sensor artifacts, reducing the need for heavy preprocessing.
  • The low Chamfer distances across overlap ratios imply the same trained model can align partial views, which is relevant for endoscopic or limited-field-of-view capture.
  • Because training is end-to-end on point cloud pairs, the architecture can be retrained on new anatomical datasets without hand-crafted features or separate optimization stages.

Reading between the lines

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

  • The equal-cardinality reconstruction of the 4DMatch target (Eq. 11) removes the partial-data setting the benchmark is designed to test; a fair check of partial-overlap robustness would run on the original targets with unequal point counts.
  • The sizable gap over the diffeomorphic baseline [13] on 4DMatch may partly reflect the reconstruction rather than the learned representation; comparing against learning-based partial-registration methods on the original benchmark would clarify this.
  • The method's gains on SynBench and ModelNet could be sensitive to the learnable affine transformation matrix providing a good initial alignment; ablating that component would reveal how much of the improvement comes from the Transformer features alone.
  • Because DeformedTissue comes from 45 pig head cadavers, the transfer to living human tissue, with its varying stiffness and perfusion, remains untested.
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 DefTransNet, an end-to-end Transformer-based network for non-rigid point cloud registration. The method combines a learnable transformation matrix, EdgeConv-based local/global feature extraction, a Transformer encoder-decoder for feature disambiguation, and a k-NN candidate displacement module regularized by Loopy Belief Propagation. The network is evaluated on four datasets (ModelNet, SynBench, 4DMatch, DeformedTissue) under varying deformation, noise, outlier, and overlap conditions, with mean distance error and Chamfer distance as metrics, against Robust-DefReg, Deep-Geo-Reg, Predator, GP-Aligner, and a diffeomorphic baseline. The central claim is that DefTransNet consistently outperforms current state-of-the-art registration networks across all datasets and conditions.

Significance. If the claims hold, DefTransNet would be a useful contribution to deformable point cloud registration, particularly for soft-tissue surgical applications. The paper provides publicly available code and data, describes the architecture in sufficient detail for reimplementation, and evaluates on both synthetic and real-world datasets. The combination of a Transformer-based descriptor with a graph-regularized LBP displacement module is a reasonable design. However, the evaluation as presented does not fully support the state-of-the-art claim: there are internal inconsistencies in the reported numbers, the only partial-overlap benchmark is modified in a way that removes the partial-overlap property, and no uncertainty quantification or statistical tests are provided. These issues are fixable but currently undermine the headline claim.

major comments (4)
  1. [Section 5.3, Table 3; Section 5.2, Table 2] The text contradicts the tables in load-bearing ways. In Section 5.3, the paper states that on ModelNet at noise 0.01 DefTransNet achieves 0.01105, 'slightly better than Robust-DefReg's 0.01075', but the table shows Robust-DefReg is better (0.01075 < 0.01105). Similarly, at noise 0.05 the table lists DefTransNet 0.0362 and Robust-DefReg 0.03440, again favoring Robust-DefReg. This directly contradicts the abstract and Section 5.2's claim of consistent SOTA outperformance. Additionally, Section 5.2 says at ModelNet deformation level 0.3 DefTransNet achieves 0.00078, but Table 2 shows 0.00409 for that level (0.00078 is the value at level 0.1). The authors must correct these mismatches and add error bars or statistical tests, as the differences are small (e.g., 0.01105 vs 0.01075).
  2. [Section 5.1, Eq. (11)] The 4DMatch benchmark is modified by generating a new target cloud Y' = tT + (X + D).RT and discarding the original target, so that source and target have equal cardinality. The paper explicitly acknowledges this is needed because the method requires equal cardinality. As a result, the partial-overlap property that defines 4DMatch/4DLoMatch is removed, so Table 5 cannot support the claimed robustness to partial data. The comparison in Table 5 also includes only the diffeomorphic baseline [13], not the learning-based methods used in the other tables, so the claim that DefTransNet outperforms state-of-the-art on overlap scenarios is unsupported. The authors should either evaluate on the original benchmark (e.g., by adapting the method to unequal cardinality through a principled subsampling scheme) or clearly frame the results as being on a modified, equal-cardinality version of the dataset and justify the modification in terms of the application context.
  3. [Tables 2-5] All reported results are point estimates without standard deviations, confidence intervals, or significance tests. Many comparisons involve differences on the order of 1e-3 or smaller (e.g., Table 3 ModelNet noise 0.01: DefTransNet 0.01105 vs Robust-DefReg 0.01075; Table 2 ModelNet deformation 0.1: both 0.00078). Without uncertainty quantification, the claimed consistent superiority cannot be assessed, and the internal contradictions noted above could plausibly stem from run-to-run variance. The authors should report multi-run statistics and at least pairwise tests where the claimed advantage is small.
  4. [Section 5.4, Table 5] The 4DMatch evaluation lacks any learning-based state-of-the-art baseline. Methods such as Lepard [31] and the Neural Deformation Pyramid [32] were specifically designed for the 4DMatch benchmark; the paper cites both but does not compare against them. The only baselines in Table 5 are the two variants of the diffeomorphic method [13]. Consequently, the statement that DefTransNet 'consistently achieves the lowest errors' across overlap ratios does not substantiate a state-of-the-art claim. The authors should add comparisons with at least one learning-based 4DMatch method, or alternatively justify why such a comparison is not possible given the modified benchmark.
minor comments (6)
  1. [Abstract] The URL in the abstract contains 'https://https://github.com/...'; this should be corrected to a single 'https://'.
  2. [Throughout] There are numerous typographical errors and inconsistent spacing, e.g., 'ef fect', 'e ffectively', and irregular equation formatting. A thorough proofread is needed.
  3. [Section 5.1] The dataset split is described as '47,738 training points, 6,400 validation points, and a test set consisting of 10,327 4DMatch points and 4,590 4DLoMatch points.' These should be 'point cloud pairs' or 'samples', not 'points'.
  4. [Section 5.1, Eq. (11)] The notation in Eq. (11), Y = tT + (X + D).RT, is ambiguous: 'tT' and 'RT' appear to mix a translation vector and rotation matrix with a transpose operator, but the meaning is not defined. Please clarify the notation and check consistency with the original reference [32].
  5. [Section 2, Related Work] The related work discusses many methods that are not compared in the experiments (e.g., Lepard, CoFiNet, Neural Deformation Pyramid). Adding a sentence on why these were not selected as baselines would help the reader understand the scope of the comparison.
  6. [Figure 2] The caption says the dataset is 'available upon request to readers', but the abstract and data link refer to the SynBench repository. Please clarify the actual availability of the DeformedTissue dataset.

Circularity Check

1 steps flagged · score 4.0 of 10

No equation-level circularity in the network itself, but the 4DMatch partial-overlap evaluation is self-referential: the target is reconstructed from the source plus the ground-truth deformation, so the tested overlap property is removed by construction.

  1. other [Section 5.1, Eq. (11); Table 5; Section 5.4]
    "For our work, the dataset is modified to generate a new target point cloud Y′ using the formula from [32], and the original target point cloud is discarded. Y = tT + (X + D).RT (11) ... by ensuring the source and target datasets have an equal number of points, we address a key limitation of our method requiring equal cardinality, making this modification both practical and well-justified."

    Table 5 is presented as evidence that DefTransNet is robust to low overlap, with overlap ratios down to 0.1. But Eq. (11) constructs Y' from the source X and the ground-truth displacement D (plus rotation and translation), and the original partial target is discarded. The transformation is applied only to points that have corresponding matches, so every source point receives a generated target point; the actual overlap in the modified test inputs is therefore 100% by construction, regardless of the original dataset's overlap label. The reported 'overlap ratio' bins are inherited from the discarded original target and no longer describe the clouds actually registered.

full rationale

The core derivation is not circular: DefTransNet maps source and target point clouds to a displacement field through a learnable transformation matrix, EdgeConv features, and Transformer attention, and no target quantity is embedded in the architecture or algebraically identified with the reported error metric. The main self-referential elements are author-created benchmarks (SynBench from SimTool, and DeformedTissue collected by the authors) and the primary baseline Robust-DefReg, which is the authors' previous work; these are generalizability and validity concerns rather than equation-level circularity because the comparisons are externally measurable and ModelNet provides an independent dataset. The 4DMatch modification is the one place where a claimed evaluation property, partial overlap, is defined out of the test by construction, and it is flagged above. Separately, Table 3's ModelNet noise numbers contradict the text: DefTransNet's 0.01105 is not better than Robust-DefReg's 0.01075, and at 0.05 noise Robust-DefReg (0.03440) beats DefTransNet (0.0362); this is a correctness risk in the SOTA claim, not a circularity. Overall, the central network derivation remains independent, so the circularity score is moderate rather than high.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on no newly invented physical entities. The key free parameters (alpha, k, transformer depth) are undisclosed hand-chosen hyperparameters. The most load-bearing assumptions are the equal-cardinality constraint (which forces benchmark modification) and the validity of the author-created deformation datasets.

free parameters (3)
  • alpha (LBP regularization weight)
    Introduced in Eq. (10) as a hyperparameter controlling the influence of the pairwise regularization term. Its value is not reported in the paper.
  • k (number of nearest neighbor candidates)
    Used in the displacement network to select candidate points for each source point. No value is given.
  • Transformer architecture hyperparameters
    Number of encoder/decoder blocks, attention heads, and feature dimensions are not specified, though they are central to the method's capacity.
assumptions (4)
  • domain assumption Source and target point clouds must have equal cardinality.
    Stated in Section 5.1: the authors modify 4DMatch to ensure equal point counts, explicitly calling it a 'key limitation of our method requiring equal cardinality'.
  • domain assumption The 4DMatch target reconstructed via Y' = tT + (X + D).RT is a faithful proxy for real deformable registration.
    Section 5.1 uses Eq. (11) to generate new targets and discards the original ones; the paper assumes this preserves the integrity of correspondences.
  • domain assumption Loopy belief propagation converges to a useful set of displacement assignments.
    Eq. (10) iterates message updates; the paper relies on LBP as a probabilistic inference step without analyzing convergence or optimality guarantees.
  • domain assumption Euclidean distance in feature space is a valid cost for correspondence selection.
    Eq. (8) defines dp_i = ||f(x_i) - f(cp_i)||^2_2, assuming that feature similarity is a reliable indicator of true correspondence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DefTransNet: A Transformer-based Method for Non-Rigid Point Cloud Registration in the Simulation of Soft Tissue Deformation." pith.science (2026). https://pith.science/paper/GG3FB3CJ

@misc{pith2026250206336,
  author       = {Pith},
  title        = {Pith review of: DefTransNet: A Transformer-based Method for Non-Rigid Point Cloud Registration in the Simulation of Soft Tissue Deformation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GG3FB3CJ}},
  note         = {Machine review of arXiv:2502.06336}
}
read the original abstract

Soft-tissue surgeries, such as tumor resections, are complicated by tissue deformations that can obscure the accurate location and shape of tissues. By representing tissue surfaces as point clouds and applying non-rigid point cloud registration (PCR) methods, surgeons can better understand tissue deformations before, during, and after surgery. Existing non-rigid PCR methods, such as feature-based approaches, struggle with robustness against challenges like noise, outliers, partial data, and large deformations, making accurate point correspondence difficult. Although learning-based PCR methods, particularly Transformer-based approaches, have recently shown promise due to their attention mechanisms for capturing interactions, their robustness remains limited in challenging scenarios. In this paper, we present DefTransNet, a novel end-to-end Transformer-based architecture for non-rigid PCR. DefTransNet is designed to address the key challenges of deformable registration, including large deformations, outliers, noise, and partial data, by inputting source and target point clouds and outputting displacement vector fields. The proposed method incorporates a learnable transformation matrix to enhance robustness to affine transformations, integrates global and local geometric information, and captures long-range dependencies among points using Transformers. We validate our approach on four datasets: ModelNet, SynBench, 4DMatch, and DeformedTissue, using both synthetic and real-world data to demonstrate the generalization of our proposed method. Experimental results demonstrate that DefTransNet outperforms current state-of-the-art registration networks across various challenging conditions. Our code and data are publicly available.

Figures

Figures reproduced from arXiv: 2502.06336 by the authors.

Figure 1
Figure 1. The proposed network architecture of DefTransNet [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 3
Figure 3. Visualization of non-rigid point cloud registration results on ModelNet dataset and under varying conditions: (a) rotation of 0.7 radians [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Visualization of non-rigid point cloud registration results on the DeformedTissue dataset under increasing deformation levels: (a) 0.3, [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Visualization of non-rigid point cloud registration results on the 4DMatch dataset under varying rotation and overlap conditions: (a) [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

61 extracted references · 55 canonical work pages

  1. [13]

    Croquet, D

    B. Croquet, D. Christiaens, S. M. Weinberg, M. Bronstein, D. Vander- meulen, and P. Claes. Unsupervised di ffeomorphic surface registration and non-linear modelling. In Medical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Conference, Stras- bourg, France, September 27–October 1, 2021, Proceedings, Part IV 24, pages 11...

    work page 2021

  2. [31]

    Li and T

    Y . Li and T. Harada. Lepard: Learning partial point cloud matching in rigid and deformable scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5554–5564, 2022

    work page 2022

  3. [32]

    Li and T

    Y . Li and T. Harada. Non-rigid point cloud registration with neural defor- mation pyramid. Advances in Neural Information Processing Systems, 35: 27757–27768, 2022

    work page 2022

  4. [1]

    Anguelov, P

    D. Anguelov, P. Srinivasan, D. Koller, S. Thrun, J. Rodgers, and J. Davis. Scape: shape completion and animation of people. ACM SIGGRAPH 2005 Papers, pages 408–416, 2005

    work page 2005

  5. [2]

    S. Ao, Q. Hu, B. Yang, A. Markham, and Y . Guo. Spinnet: Learning a general surface descriptor for 3d point cloud registration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11753–11762, 2021

    work page 2021

  6. [3]

    Y . Aoki, H. Goforth, R. A. Srivatsan, and S. Lucey. Pointnetlk: Robust & efficient point cloud registration using pointnet. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7163–7172, 2019

    work page 2019

  7. [4]

    Bauer, T

    D. Bauer, T. Patten, and M. Vincze. Reagent: Point cloud registration using imitation and reinforcement learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14586– 14594, 2021

    work page 2021

  8. [5]

    Belhasin, G

    O. Belhasin, G. Bar-Shalom, and R. El-Yaniv. Transboost: improving the best imagenet performance using deep transduction. Advances in Neural Information Processing Systems, 35:28363–28373, 2022

    work page 2022

Show all 61 references
  1. [6]

    P. J. Besl and N. D. McKay. Method for registration of 3-d shapes. In Sensor fusion IV: control paradigms and data structures, volume 1611, pages 586–606. Spie, 1992

    1992

  2. [7]

    F. Bogo, J. Romero, M. Loper, and M. J. Black. Faust: Dataset and evaluation for 3d mesh registration. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 3794–3801, 2014

    2014

  3. [8]

    Bozic, M

    A. Bozic, M. Zollhofer, C. Theobalt, and M. Nießner. Deepdeform: Learn- ing non-rigid rgb-d reconstruction with semi-supervised data. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7002–7012, 2020

    2020

  4. [9]

    Castellani and A

    U. Castellani and A. Bartoli. 3d shape registration. In3D Imaging, Analysis and Applications, pages 353–411. Springer, 2020

    2020

  5. [10]

    Chalvidal, T

    M. Chalvidal, T. Serre, and R. VanRullen. Learning functional transduction. Advances in Neural Information Processing Systems, 36, 2024

    2024

  6. [11]

    A. X. Chang, T. Funkhouser, L. Guibas, P. Hanrahan, Q. Huang, Z. Li, S. Savarese, M. Savva, S. Song, H. Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015

    2015 arXiv

  7. [12]

    X. Chen, J. Luo, Y . Ren, T. Cui, and M. Zhang. Mafnet: a two-stage multi- ple attention fusion network for partial-to-partial point cloud registration. Measurement Science and Technology, 35(12):125113, 2024

    2024

  8. [14]

    B. Deng, Y . Yao, R. M. Dyke, and J. Zhang. A survey of non-rigid 3d registration. Computer Graphics Forum, 41(2):559–589, 2022

    2022

  9. [15]

    Devlin, M.-W

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018

    2018 arXiv

  10. [16]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Un- terthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020

    2010 arXiv

  11. [17]

    Geiger, P

    A. Geiger, P. Lenz, C. Stiller, and R. Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11): 1231–1237, 2013

    2013

  12. [18]

    Grosse, M

    R. Grosse, M. K. Johnson, E. H. Adelson, and W. T. Freeman. Ground truth dataset and baseline evaluations for intrinsic image algorithms. In 2009 IEEE 12th International Conference on Computer Vision , pages 2335–2342. IEEE, 2009

    2009

  13. [19]

    Guo, J.-X

    M.-H. Guo, J.-X. Cai, Z.-N. Liu, T.-J. Mu, R. R. Martin, and S.-M. Hu. Pct: Point cloud transformer. Computational Visual Media, 7(2):187–199, 2021

    2021

  14. [20]

    Hansen and M

    L. Hansen and M. P. Heinrich. Deep learning based geometric registration for medical images: How accurate can we get without visual features? In Information Processing in Medical Imaging: 27th International Confer- ence, IPMI 2021, Virtual Event, June 28–June 30, 2021, Proceed...

    2021

  15. [21]

    Hansen, D

    L. Hansen, D. Dittmer, and M. P. Heinrich. Learning deformable point set registration with regularized dynamic graph cnns for large lung motion in copd patients. In International Workshop on Graph Learning in Medical Imaging, pages 53–61. Springer, 2019

    2019

  16. [22]

    M. Horn, N. Engel, V . Belagiannis, M. Buchholz, and K. Dietmayer. Deepclr: Correspondence-less architecture for deep end-to-end point cloud registration. In 2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC), pages 1–7. IEEE, 2020. ISBN 978-1...

    2020

  17. [23]

    X. Hu, D. Zhang, J. Chen, Y . Wu, and Y . Chen. Nrtnet: An unsupervised method for 3d non-rigid point cloud registration based on transformer. Sensors, 22(14):5128, 2022

    2022

  18. [24]

    Huang, Z

    S. Huang, Z. Gojcic, M. Usvyatsov, A. Wieser, and K. Schindler. Predator: Registration of 3d point clouds with low overlap. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 4267–4276, 2021

    2021

  19. [25]

    Huang, G

    X. Huang, G. Mei, J. Zhang, and R. Abbas. A comprehensive survey on point cloud registration. arXiv preprint arXiv:2103.02690, 2021

    2021 arXiv

  20. [26]

    Jaderberg, K

    M. Jaderberg, K. Simonyan, A. Zisserman, et al. Spatial transformer networks. Advances in neural information processing systems, 28, 2015

    2015

  21. [27]

    Jiang, M

    P. Jiang, M. Sun, and R. Huang. Neural intrinsic embedding for non-rigid point cloud matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21835–21845, 2023

    2023

  22. [28]

    Jiang, J

    X. Jiang, J. Ma, G. Xiao, Z. Shao, and X. Guo. A review of multimodal image matching: Methods and applications. Information Fusion, 73:22–71, 2021

    2021

  23. [29]

    M. Kinz. SimTool-SynBench, 2023. URL https://doi.org/10.115 88/data/R9IKCF

    2023

  24. [30]

    X. Li, J. K. Pontes, and S. Lucey. Pointnetlk revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12763–12772, 2021

    2021

  25. [33]

    Y . Li, H. Takehara, T. Taketomi, B. Zheng, and M. Nießner. 4dcomplete: Non-rigid motion estimation beyond the observable surface. In Proceed- ings of the IEEE/CVF International Conference on Computer Vision, pages 12706–12716, 2021

    2021

  26. [34]

    Mahmood and S

    B. Mahmood and S. Han. 3d registration of indoor point clouds for augmented reality. In Computing in Civil Engineering 2019: Visualization, Information Modeling, and Simulation, pages 1–8. American Society of Civil Engineers Reston, V A, 2019

    2019

  27. [35]

    M¨annle, J

    D. M¨annle, J. Pohlmann, S. Monji-Azad, J. Hesser, N. Rotter, A. Affolter, A. Lammert, B. Kramer, S. Ludwig, L. Huber, et al. Artificial intelligence directed development of a digital twin to measure soft tissue shift during head and neck surgery. Plos one, 18(8):e0287081, 2023

    2023

  28. [36]

    Melzi, R

    S. Melzi, R. Marin, E. Rodol `a, U. Castellani, J. Ren, A. Poulenard, P. Wonka, and M. Ovsjanikov. Shrec 2019: Matching humans with dif- ferent connectivity. In Eurographics Workshop on 3D Object Retrieval, volume 7, page 3. The Eurographics Association, 2019

    2019

  29. [37]

    T. Min, C. Song, E. Kim, and I. Shim. Distinctiveness oriented positional equilibrium for point cloud registration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5490–5498, 2021

    2021

  30. [38]

    Monji-Azad, J

    S. Monji-Azad, J. Hesser, and N. L ¨ow. A review of non-rigid transfor- mations and learning-based 3d point cloud registration methods. ISPRS Journal of Photogrammetry and Remote Sensing, 196:58–72, 2023

    2023

  31. [39]

    Monji-Azad, M

    S. Monji-Azad, M. Kinz, J. Hesser, and N. L ¨ow. Simtool: A toolset for soft body simulation using flex and unreal engine. Software Impacts, 17: 100521, 2023

    2023

  32. [40]

    Monji-Azad, M

    S. Monji-Azad, M. Kinz, D. M ¨annel, C. Scherl, and J. Hesser. Robust- defreg: a robust coarse to fine non-rigid point cloud registration method based on graph convolutional neural networks. Measurement Science and Technology, 36(1):015426, 2024

    2024

  33. [41]

    Monji-Azad, D

    S. Monji-Azad, D. M¨annle, J. Hesser, J. Pohlmann, N. Rotter, A. Affolter, C. A. Weis, S. Ludwig, and C. Scherl. Point cloud registration for measur- ing shape dependence of soft tissue deformation by digital twins in head and neck surgery. Biomedicine hub, 9(1):9–15, 2024

    2024

  34. [42]

    Pons-Moll, S

    G. Pons-Moll, S. Pujades, S. Hu, and M. J. Black. Clothcap: Seamless 4d clothing capture and retargeting. ACM Transactions on Graphics (ToG), 36(4):1–15, 2017

    2017

  35. [43]

    Z. Qin, H. Yu, C. Wang, Y . Guo, Y . Peng, and K. Xu. Geometric trans- former for fast and robust point cloud registration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11143–11152, 2022

    2022

  36. [44]

    Z. Qin, H. Yu, C. Wang, Y . Peng, and K. Xu. Deep graph-based spatial consistency for robust non-rigid point cloud registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5394–5403, 2023

    2023

  37. [45]

    Shotton, B

    J. Shotton, B. Glocker, C. Zach, S. Izadi, A. Criminisi, and A. Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb-d images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2930–2937, 2013

    2013

  38. [46]

    L. Sun, Z. Zhang, R. Zhong, D. Chen, L. Zhang, L. Zhu, Q. Wang, G. Wang, J. Zou, and Y . Wang. A weakly supervised graph deep learning framework for point cloud registration. IEEE Transactions on Geoscience and Remote Sensing, 60:1–12, 2022

    2022

  39. [47]

    Varol, J

    G. Varol, J. Romero, X. Martin, N. Mahmood, M. J. Black, I. Laptev, and C. Schmid. Learning from synthetic humans. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 109–117, 2017

    2017

  40. [48]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017

    2017

  41. [49]

    L. Wang, N. Zhou, H. Huang, J. Wang, X. Li, and Y . Fang. Gp-aligner: Un- supervised groupwise nonrigid point set registration based on optimizable group latent descriptor. IEEE Transactions on Geoscience and Remote Sensing, 60:1–13, 2022

    2022

  42. [50]

    Wang and M.-K

    Q. Wang and M.-K. Kim. Applications of 3d point cloud data in the construction industry: A fifteen-year review from 2004 to 2018. Advanced Engineering Informatics, 39:306–319, 2019

    2004

  43. [51]

    R. Wang, J. Peethambaran, and D. Chen. Lidar point clouds to 3-d urban models : a review. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 11(2):606–627, 2018

    2018

  44. [52]

    Wang and J

    Y . Wang and J. M. Solomon. Deep closest point: Learning representations for point cloud registration. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3523–3532, 2019

    2019

  45. [53]

    Wang and J

    Y . Wang and J. M. Solomon. Prnet: Self-supervised learning for partial-to- partial registration. Advances in neural information processing systems, 32, 2019

    2019

  46. [54]

    Y . Wang, Y . Sun, Z. Liu, S. E. Sarma, M. M. Bronstein, and J. M. Solomon. Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog), 38(5):1–12, 2019

    2019

  47. [55]

    Z. Wu, S. Song, A. Khosla, F. Yu, L. Zhang, X. Tang, and J. Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 14 1912–1920, 2015

    1912

  48. [56]

    F. Yang, L. Guo, Z. Chen, and W. Tao. One-inlier is first: Towards efficient position encoding for point cloud registration. Advances in Neural Information Processing Systems, 35:6982–6995, 2022

    2022

  49. [57]

    Z. J. Yew and G. H. Lee. Regtr: End-to-end point cloud correspondences with transformers. In Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 6677–6686, 2022

    2022

  50. [58]

    H. Yu, F. Li, M. Saleh, B. Busam, and S. Ilic. Cofinet: Reliable coarse- to-fine correspondences for robust pointcloud registration. Advances in Neural Information Processing Systems, 34:23872–23884, 2021

    2021

  51. [59]

    A. Zeng, S. Song, M. Nießner, M. Fisher, J. Xiao, and T. Funkhouser. 3dmatch: Learning local geometric descriptors from rgb-d reconstructions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1802–1811, 2017

    2017

  52. [60]

    Zhang, H

    S. Zhang, H. Tong, J. Xu, and R. Maciejewski. Graph convolutional networks: a comprehensive review. Computational Social Networks, 6(1): 1–23, 2019

    2019

  53. [61]

    H. Zhao, L. Jiang, J. Jia, P. H. Torr, and V . Koltun. Point transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16259–16268, 2021. 15

    2021

Pith tools

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