Pith. sign in

REVIEW 4 major objections 5 minor 32 references

GeoCD: A Differential Local Approximation for Geodesic Chamfer Distance

T0 review · 4 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 GeoCD replaces Euclidean Chamfer distance with a differentiable geodesic-like distance computed on a multi-hop kNN graph, and one epoch of fine-tuning with it improves point cloud reconstruction across architectures and datasets.

desk verdict Novel differentiable geodesic loss, but the experiments lack a standard-CD fine-tuning control, so the headline improvement claim isn't yet established. read the letter →

arxiv 2506.23478 v1 pith:S4VKP5CF submitted 2025-06-30 cs.CV

Pedro Alonso , Tianrui Li , Chongshou Li This is my paper
classification cs.CV
keywords GeoCDgeodesicdistancepointcloudreconstructionChamferk-nearest-neighborgraphmin-plusmatrixmultiplicationsoftminfine-tuning
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

GeoCD is a loss function that approximates geodesic distance between two point clouds by building a multi-hop k-nearest-neighbor graph on the merged point sets and propagating distances along graph paths. The paper argues that standard Chamfer Distance, which measures Euclidean distances to nearest neighbors, misses the intrinsic geometry of curved surfaces, and that replacing it with GeoCD during fine-tuning improves reconstruction quality. Experiments across two architectures (a simple autoencoder and Point Transformer v3) and two datasets (ModelNet40 and ShapeNetPart) show consistent gains in Chamfer Distance, Hausdorff Distance, and F1 score after only a single epoch of fine-tuning. The authors position GeoCD as a practical, fully differentiable alternative to graph-based geodesic methods such as Dijkstra or Fast Marching, which are non-differentiable and costly.

What carries the argument

The central object is the multi-hop k-nearest-neighbor graph built on the concatenated predicted and ground-truth point clouds. Its edges are weighted by Euclidean distances, repeated min-plus matrix multiplication computes shortest paths of up to two hops, and a softmin operation converts the resulting distance matrix into a smooth, differentiable loss. What this does is let two points that are close on the surface but far in Euclidean space connect through intermediate points, so the loss reflects surface geometry rather than straight-line distance.

What would settle it

Take a point cloud of a known mesh, compute GeoCD between the original and a version where points are shifted tangentially along the surface so Euclidean distances change but geodesic distances do not; if GeoCD does not track the geodesic distance better than Euclidean CD does, the graph approximation is not capturing the surface topology. Alternatively, if the reported gains disappear when the masking threshold is changed or when k or the hop count is varied, the effect is tied to the specific graph configuration rather than to geodesic geometry.

Watch

Extended reading notes

Core claim

The paper's central claim is that a topology-aware, fully differentiable approximation of geodesic distance can be computed as a multi-hop shortest-path problem on a k-nearest-neighbor graph over the union of predicted and ground-truth points, and that using this distance in place of Euclidean nearest-neighbor matching yields consistently better reconstructions. Concretely, the method constructs the graph, propagates distances via min-plus matrix multiplication for two hops, softmin-aggregates the resulting cross-set distances, and uses the result as a loss. Fine-tuning a model that was already trained with Chamfer Distance for a single epoch with this loss lowers Chamfer Distance and Hausdorff Distance and raises F1@1% on both ModelNet40 and ShapeNetPart for both the autoencoder and PTv3 backbones.

Load-bearing premise

The k-nearest-neighbor graph built on the aligned predicted and ground-truth points truly captures the shape's surface, so that a two-hop path on that graph approximates the true geodesic curve between the two clouds.

Editorial extensions

If this is right

  • A single epoch of GeoCD fine-tuning improves reconstruction quality across different architectures and datasets, so it can serve as a drop-in refinement for any model already trained with standard Chamfer Distance.
  • Because all three evaluation metrics improve together, the gain reflects genuine geometric alignment rather than a trade-off between measures.
  • Being fully differentiable, GeoCD can be used inside deep learning pipelines where traditional geodesic algorithms such as Dijkstra or Fast Marching cannot.
  • The computational cost of GeoCD depends only weakly on model complexity, so its relative overhead is smaller for large models like PTv3 than for small autoencoders.

Reading between the lines

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

  • Extending the graph with more hops or a connectivity estimate that is not purely Euclidean could make GeoCD applicable earlier in training, not only as a final fine-tuning step.
  • Because the paper observes saturation after one epoch, the main benefit is likely a local correction of surface geometry; using GeoCD as the primary loss throughout training might yield larger gains but would require solving the alignment problem the paper identifies.
  • The same graph construction could be used to define other geodesic-aware losses, such as a geodesic Earth Mover's Distance, for tasks like shape completion or generation.
  • A direct comparison of GeoCD against HyperCD or DCD on the same backbones would clarify whether the improvement comes from geodesic awareness specifically or from any refinement of the Chamfer loss.
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 / 5 minor

Summary. The paper proposes GeoCD, a differentiable approximation of geodesic Chamfer distance for 3D point cloud reconstruction. The method builds a multi-hop k-nearest-neighbor graph on the merged predicted and ground truth point clouds, propagates distances through the graph, and uses a softmin operation to define a loss. The authors fine-tune models pre-trained with standard Chamfer Distance for one epoch on ModelNet40 and ShapeNetPart, reporting consistent improvements in CD, Hausdorff distance, and F1@1% across two architectures. Ablations study the graph connectivity parameter k and the number of fine-tuning epochs, and the paper concludes with a discussion of limitations and computational cost.

Significance. The paper's idea is practically appealing: a simple, differentiable loss that approximates geodesic structure could improve point-cloud reconstruction during fine-tuning. The formulation is clear, the softmin over the multi-hop distance matrix is a natural way to preserve differentiability, and the authors provide ablations on k, fine-tuning duration, computational time, and an honest list of limitations. If the empirical claim were properly controlled, this could be a useful contribution to the point-cloud learning literature. As it stands, however, the main evidence does not isolate the effect of the geodesic approximation from the effect of additional training, so the significance of the paper is conditional on additional experiments.

major comments (4)
  1. [§4.2 / Table 1] The central comparison in Table 1 is between a model that has received no additional optimization and the same model fine-tuned with GeoCD for one epoch. This conflates the effect of the geodesic loss with the effect of one additional training epoch; a model fine-tuned with the standard Chamfer Distance under identical settings (same optimizer, learning rate, batch size, epoch count) is never run. Section 4.3.2 varies only the number of GeoCD fine-tuning epochs and never includes a CD fine-tuning arm, so that ablation does not resolve the confound. Please add a CD fine-tuning control for every model/dataset pair and report paired differences; without it, the improvements in CD, HD, and F1@1% cannot be attributed to GeoCD.
  2. [§4.1.3] The masking strategy is described only as masking a point once it is 'within a certain distance threshold' from any point in the other point cloud, but no threshold value or normalization is given. This threshold is a free parameter used in all experiments and must be specified for reproducibility. More importantly, the justification that masking cannot affect the result because any longer path would only increase the distance is incomplete: a masked point can serve as an intermediate node on a path that is shorter than any path avoiding that point, so removing it can increase the computed shortest-path distance. Please provide either a formal argument or an ablation that measures the effect of the masking threshold on the computed distances and on the reported metrics.
  3. [§4.1.3 / Table 1] All quantitative claims rest on single runs with no error bars or multiple seeds. The reported differences are small relative to the precision of the tables (e.g., 103×CD changes of 0.04–0.10 and HD changes of 0.06–0.56), so without variance estimates it is impossible to judge whether the improvements are significant or within run-to-run noise. Please report mean and standard deviation over at least three seeds, and ideally paired significance tests, for both the GeoCD and CD fine-tuning arms.
  4. [§3.1, Eq. (2)] The statement that after normalization to a unit bounding box 'all true pairwise distances are smaller than 1' is incorrect: the maximum Euclidean distance between two points in the unit cube is sqrt(3) ≈ 1.73. Because non-neighbor entries are set to 1 instead of infinity, any actual pair distance greater than 1 is represented by a smaller value, and exp(-1) in the softmin (Eq. 4) can then exceed exp(-d_true) for true distances above 1, distorting the loss. The normalization or the default value for non-neighbor entries needs to be corrected, for example by using the actual maximum distance or setting a value provably larger than all true distances.
minor comments (5)
  1. [§4.3.1] The hyperparameters k and the number of hops are selected on ModelNet40 and then applied to ShapeNetPart; this is a mild selection issue. A brief sensitivity analysis on ShapeNetPart, or an explicit justification for transferring the selected values, would strengthen the cross-dataset claim.
  2. [§4.4.2] The runtime comparison is phrased ambiguously: 'the additional computational cost per epoch is comparable only for PTv3 to the total training time of CD' should be rewritten to clarify that for PTv3 one GeoCD epoch costs roughly as much as the entire CD training run, while for the autoencoder it is much more expensive than CD training.
  3. [§1] The sentence 'comparing reconstructed and ground truth shapes is a non-trivial' is missing a noun; it should read 'is a non-trivial problem' or similar.
  4. [§4.1.4, Eq. (6)] In the Hausdorff distance definition, '∥ · ∥denotes' should have a space before 'denotes'.
  5. [Table 1] The table headings '103×CD' and '102×HD' should be defined in the table caption as 10^3 × CD and 10^2 × HD to avoid reader confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: GeoCD is an independent heuristic loss; the main weakness is a missing fine-tuning control, which is an experimental confound rather than a circular derivation.

full rationale

GeoCD is constructed from first principles as a differentiable multi-hop kNN graph distance (Eqs. 2-5). The loss is defined directly on merged predicted and ground-truth point sets via min-plus propagation and softmin; it does not take Chamfer Distance as an input, nor is any parameter fitted to the reported evaluation metrics. Hyperparameter k is selected by ablation on ModelNet40 and then applied to both datasets, which is a mild selection issue, not a circular reduction. There are no load-bearing self-citations: the references are to external prior work (PointNet, PTv3, HyperCD, etc.), and no uniqueness theorem is invoked. The paper honestly acknowledges in Section 4.4.1 and Section 5 that GeoCD requires a coarse alignment and is therefore used only for fine-tuning, which is a scope limitation rather than a circular step. The most substantive concern is experimental: Table 1 compares GeoCD fine-tuned models to baselines with no fine-tuning, so the reported gains could partly stem from additional optimization rather than geodesic geometry; however, this is a missing-control/correctness issue, not a case where a prediction reduces to its own inputs by construction. No equation or argument in the paper makes the central claim equivalent to a fitted parameter or a self-citation chain.

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

GeoCD is a heuristic loss. The main free parameters are the graph neighborhood size k, the number of hops, and the masking threshold. The central assumptions are that the kNN graph on the merged clouds approximates surface geodesics and that masking close points does not alter shortest paths.

free parameters (3)
  • k (number of nearest neighbors) = 5
    Chosen by ablation on ModelNet40 (k=3,5,10); k=5 balances accuracy and runtime (Section 4.3.1).
  • number of hops n = 2
    Set to 2 in all experiments; no ablation is reported for the number of hops (Section 4.1.3).
  • masking distance threshold = unspecified
    Section 4.1.3 mentions 'a certain distance threshold' for masking points in multi-hop propagation but never states the value.
assumptions (4)
  • domain assumption The kNN graph over the merged point set approximates the true surface geodesics.
    The method's core premise; if the graph does not follow the surface (thin structures, holes, misalignment), the computed distances are not geodesic. Acknowledged in Section 4.4.1.
  • domain assumption Point clouds are normalized to a unit bounding box so all pairwise distances are below 1, justifying setting non-neighbor entries to 1.
    Section 3.1; only valid for normalized inputs.
  • standard math Softmin with negative log-sum-exp approximates the minimum sufficiently well for gradient-based optimization.
    Section 3.2; standard smooth approximation.
  • ad hoc to paper Masking points within a threshold from the other cloud does not change the shortest paths that matter.
    Section 4.1.3 claims this, but the argument only covers paths whose length increases, not paths that use a masked point as an intermediate.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GeoCD: A Differential Local Approximation for Geodesic Chamfer Distance." pith.science (2026). https://pith.science/paper/S4VKP5CF

@misc{pith2026250623478,
  author       = {Pith},
  title        = {Pith review of: GeoCD: A Differential Local Approximation for Geodesic Chamfer Distance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S4VKP5CF}},
  note         = {Machine review of arXiv:2506.23478}
}
read the original abstract

Chamfer Distance (CD) is a widely adopted metric in 3D point cloud learning due to its simplicity and efficiency. However, it suffers from a fundamental limitation: it relies solely on Euclidean distances, which often fail to capture the intrinsic geometry of 3D shapes. To address this limitation, we propose GeoCD, a topology-aware and fully differentiable approximation of geodesic distance designed to serve as a metric for 3D point cloud learning. Our experiments show that GeoCD consistently improves reconstruction quality over standard CD across various architectures and datasets. We demonstrate this by fine-tuning several models, initially trained with standard CD, using GeoCD. Remarkably, fine-tuning for a single epoch with GeoCD yields significant gains across multiple evaluation metrics.

Figures

Figures reproduced from arXiv: 2506.23478 by the authors.

Figure 1
Figure 1. Illustration of the geodesic distance approximation. (a) Euclidean distance between two points. (b) kNN graph for [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Qualitative comparison of point cloud reconstructions. Each row shows a ground truth shape (left), the output of a model trained [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 29 canonical work pages

  1. [1]

    Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Mano- lis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu

    Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Mano- lis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. Shapenet: An information-rich 3d model reposi- tory, 2015. 4

    work page 2015

  2. [2]

    A point set generation network for 3d object reconstruction from a single image

    Haoqiang Fan, Hao Su, and Leonidas Guibas. A point set generation network for 3d object reconstruction from a single image. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2463–2471, 2017. 1

    work page 2017

  3. [3]

    Kinectfusion: real-time 3d recon- struction and interaction using a moving depth camera

    Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, Andrew Davison, and Andrew Fitzgibbon. Kinectfusion: real-time 3d recon- struction and interaction using a moving depth camera. In Proceedings of the 24th Annual ACM Symposium on User In- terface Software and Technology...

    work page 2011

  4. [4]

    Gal: Geometric adversarial loss for single-view 3d-object recon- struction

    Li Jiang, Shaoshuai Shi, Xiaojuan Qi, and Jiaya Jia. Gal: Geometric adversarial loss for single-view 3d-object recon- struction. In Computer Vision – ECCV 2018: 15th European Conference, Munich, Germany, September 8-14, 2018, Pro- ceedings, Part VIII, page 820–834, Berlin, Heidelberg, 2018. Springer-Verlag. 1

    work page 2018

  5. [5]

    Yuqi Jiang, Jing Li, Yanran Dai, Haidong Qin, Xiaoshi Zhou, Yong Zhang, Hongwei Liu, Kefan Yan, and Tao Yang. Rt3dhvc: A real-time human holographic video conferenc- ing system with a consumer rgb-d camera array.IEEE Trans- actions on Circuits and Systems for Video Technology, 35(5): 4226–4241, 2025. 1

    work page 2025

  6. [6]

    Adam: A method for stochastic opti- mization

    Diederik P Kingma. Adam: A method for stochastic opti- mization. arXiv preprint arXiv:1412.6980, 2014. 4

    arXiv 2014

  7. [7]

    Lang, Sourabh V ora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom

    Alex H. Lang, Sourabh V ora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. PointPillars: Fast Encoders for Object Detection from Point Clouds. arXiv e-prints, art. arXiv:1812.05784, 2018. 1

    arXiv 2018

  8. [8]

    Point Cloud GAN

    Chun-Liang Li, Manzil Zaheer, Yang Zhang, Barnabas Poc- zos, and Ruslan Salakhutdinov. Point Cloud GAN. arXiv e-prints, art. arXiv:1810.05795, 2018. 1

    work page Pith review arXiv 2018

Show all 32 references
  1. [9]

    Proxyformer: Proxy alignment assisted point cloud com- pletion with missing part sensitive transformer

    Shanshan Li, Pan Gao, Xiaoyang Tan, and Mingqiang Wei. Proxyformer: Proxy alignment assisted point cloud com- pletion with missing part sensitive transformer. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9466–9475, 2023. 1

    2023

  2. [10]

    Yamada, and Ziming Zhang

    Fangzhou Lin, Yun Yue, Songlin Hou, Xuechu Yu, Yajun Xu, Kazunori D. Yamada, and Ziming Zhang. Hyperbolic cham- fer distance for point cloud completion. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14549–14560, 2023. 2

    2023

  3. [11]

    Morphing and sampling network for dense point cloud completion

    Minghua Liu, Lu Sheng, Sheng Yang, Jing Shao, and Shimin Hu. Morphing and sampling network for dense point cloud completion. In AAAI Conference on Artificial Intelligence ,

  4. [12]

    6- DOF GraspNet: Variational Grasp Generation for Object Manipulation

    Arsalan Mousavian, Clemens Eppner, and Dieter Fox. 6- DOF GraspNet: Variational Grasp Generation for Object Manipulation . In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2901–2910, Los Alami- tos, CA, USA, 2019. IEEE Computer Society. 1

    2019

  5. [13]

    Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J

    Richard A. Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J. Davison, Push- meet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE International Symposium on Mixe...

    2011

  6. [14]

    Newcombe, Dieter Fox, and Steven M

    Richard A. Newcombe, Dieter Fox, and Steven M. Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In 2015 IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR) , pages 343– 352, 2015. 1

    2015

  7. [15]

    Qi, Hao Su, Kaichun Mo, and Leonidas J

    Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1, 2

    2017

  8. [16]

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space

    Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in Neural Informa- tion Processing Systems. Curran Associates, Inc., 2017. 1, 2

    2017

  9. [17]

    Scalability in Perception for Au- tonomous Driving: Waymo Open Dataset

    Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Sheng Zha...

    1912 arXiv

  10. [18]

    Ang, and Gim Hee Lee

    Xiaogang Wang, Marcelo H. Ang, and Gim Hee Lee. Cas- caded refinement network for point cloud completion with self-supervision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8139–8150, 2022. 1

    2022

  11. [19]

    Sarma, Michael M

    Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic 8 graph cnn for learning on point clouds. ACM Trans. Graph., 38(5), 2019. 1, 2

    2019

  12. [20]

    Density-aware chamfer distance as a com- prehensive metric for point cloud completion

    Tong Wu, Liang Pan, Junzhe Zhang, Tai Wang, Ziwei Liu, and Dahua Lin. Density-aware chamfer distance as a com- prehensive metric for point cloud completion. In Proceed- ings of the 35th International Conference on Neural Infor- mation Processing Systems, Red Hook, NY , USA, 20...

    2021

  13. [21]

    Point transformer v2: Grouped vector atten- tion and partition-based pooling

    Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Heng- shuang Zhao. Point transformer v2: Grouped vector atten- tion and partition-based pooling. Advances in Neural Infor- mation Processing Systems, 35:33330–33342, 2022. 2, 4

    2022

  14. [22]

    Point transformer v3: Simpler faster stronger

    Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xi- hui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4840–4851, 2024. 2, 4

    2024

  15. [23]

    3d shapenets: A deep representation for volumetric shapes

    Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Lin- guang Zhang, Xiaoou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 4

    2015

  16. [24]

    Snowflakenet: Point cloud completion by snowflake point deconvolution with skip-transformer

    Peng Xiang, Xin Wen, Yu-Shen Liu, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Zhizhong Han. Snowflakenet: Point cloud completion by snowflake point deconvolution with skip-transformer. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision (ICCV), pages 5499...

    2021

  17. [25]

    Second: Sparsely embed- ded convolutional detection

    Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embed- ded convolutional detection. Sensors, 18(10), 2018. 1

    2018

  18. [26]

    Pointflow: 3d point cloud generation with continuous normalizing flows

    Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4540–4549, 2019. 1

    2019

  19. [27]

    Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Shef- fer, and Leonidas Guibas

    Li Yi, Vladimir G. Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Shef- fer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Trans. Graph., 35(6), 2016. 4

    2016

  20. [28]

    Doublefu- sion: Real-time capture of human performances with inner body shapes from a single depth sensor

    Tao Yu, Jianhui Zhao, Zerong Zheng, Kaiwen Guo, Qionghai Dai, Hao Li, Gerard Pons-Moll, and Yebin Liu. Doublefu- sion: Real-time capture of human performances with inner body shapes from a single depth sensor. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(...

    2020

  21. [29]

    Pointr: Diverse point cloud comple- tion with geometry-aware transformers

    Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud comple- tion with geometry-aware transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12498–12507, 2021. 1

    2021

  22. [30]

    Adapointr: Diverse point cloud completion with adaptive geometry-aware transformers

    Xumin Yu, Yongming Rao, Ziyi Wang, Jiwen Lu, and Jie Zhou. Adapointr: Diverse point cloud completion with adaptive geometry-aware transformers. IEEE Trans. Pattern Anal. Mach. Intell., 45(12):14114–14130, 2023

    2023

  23. [31]

    Pcn: Point completion network

    Wentao Yuan, Tejas Khot, David Held, Christoph Mertz, and Martial Hebert. Pcn: Point completion network. In 2018 International Conference on 3D Vision (3DV) , pages 728– 737, 2018. 1

    2018

  24. [32]

    Point transformer

    Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259–16268, 2021. 2, 4 9

    2021

Pith tools

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