BAGNet: A Boundary-Aware Graph Attention Network for 3D Point Cloud Semantic Segmentation

Jianzong Wang; Jiguang Wan; Kai Lu; Shenglin He; Wei Tao; Xiaoyang Qu

REVIEW 5 major objections 5 minor 33 references

BAGNet: A Boundary-Aware Graph Attention Network for 3D Point Cloud Semantic Segmentation

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read BAGNet claims that spending graph attention only on boundary points yields state-of-the-art 3D point cloud semantic segmentation accuracy while cutting inference time.

desk verdict A sensible boundary-focused architecture is undermined by an unspecified boundary detector and sloppy equations; the idea is worth refereeing, not the current write-up. read the letter →

arxiv 2506.00475 v1 pith:3IE2SMCK submitted 2025-05-31 cs.CV

Wei Tao , Xiaoyang Qu , Kai Lu , Jiguang Wan , Shenglin He , Jianzong Wang This is my paper

classification cs.CV

keywords pointcloudsemanticsegmentationgraphattentionnetworkboundarypointsedgevertexfusionpooling3DdeeplearningnormalvectorsShapeNet

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

BAGNet claims that in 3D point cloud semantic segmentation you can get both higher accuracy and lower inference time by splitting points into boundary points and non-boundary points, then spending graph attention only on the boundary ones. Boundary points, where semantic regions meet, have chaotic normal-vector distributions that carry intricate spatial structure; non-boundary points are easier and can be handled by a cheap shared MLP. Across the ShapeNet part dataset and S3DIS Area 5, the paper reports the best mean intersection-over-union (86.2 and 56.5) and the lowest inference time (80.2 ms on S3DIS) among the compared methods. The practical stake is that one can keep the expressiveness of graph-based point cloud networks without paying their usual cost of building a graph for every point.

What carries the argument

The load-bearing object is the boundary-aware graph attention layer (BAGLayer), a per-boundary-point graph whose edges are weighted by $\bar{W}_{ij} = \ln(|\bar{p}_i - \bar{p}_{ij}|)$ and fused into both the central point and its neighbors before three separate MLPs encode central, neighbor, and edge features. The fused encodings are combined as $y'_i = h(\hat{p}'_i + \sum_j \hat{p}'_{ij} - \sum_j \bar{W}'_{ij})$, and a softmax over the components of $y'_i$ yields attention coefficients $a_{ij} = e^{y'_{ij}} / \sum_j e^{y'_{ij}}$ used to weight the neighbor features. This concentrates graph computation on boundary points while a pointwise MLP handles the rest, and a lightweight attention pooling layer (KNN, MLP to 512 dimensions, max pooling) supplies global context. The logarithm in the edge weight is meant to eliminate the linear relationship between distance and influence.

What would settle it

Run the identical BAGNet architecture with the boundary detector replaced by a random split of points into the same fraction of boundary and non-boundary points; if ShapeNet mIoU does not drop below 86.2 or S3DIS inference time does not rise materially above 80.2 ms, then the reported gains are not caused by the geometric boundary detection. Alternatively, re-run with all points processed by the BAGLayer: if accuracy does not improve or latency does not increase, the compute-savings premise fails.

Watch

Extended reading notes

Core claim

The central claim is that concentrating computational effort on semantic boundary points is enough to outperform state-of-the-art point cloud segmentation while reducing inference time. The paper observes that boundary points show chaotic normal-vector directions relative to their neighbors, so it partitions the cloud into boundary and non-boundary points from normal-vector statistics. Boundary points are organized into k-nearest-neighbor graphs; a boundary-aware graph attention layer (BAGLayer) fuses vertex and edge information, with edge weights $\bar{W}_{ij} = \ln(|\bar{p}_i - \bar{p}_{ij}|)$, and computes normalized attention coefficients per boundary point to aggregate neighbor features. Non-boundary points pass through a pointwise MLP, and a lightweight attention pooling layer extracts global features from the raw cloud, with all features concatenated for a shared-MLP classifier. Reported results: 86.2 mIoU on ShapeNet and 56.5 mIoU on S3DIS Area 5 with 80.2 ms inference, beating the listed baselines on both accuracy and speed, with ablation attributing roughly 3.9% mIoU to the BAGLayer, 2.2% to the attention pooling layer, and 6.1% combined.

Load-bearing premise

The method depends on a reliable automatic way to tell boundary points from non-boundary points using normal-vector statistics, but the paper does not specify the algorithm, threshold, or validation of that detector.

Editorial extensions

If this is right

On ShapeNet, BAGNet raises average mIoU to 86.2, with the largest per-category gains on objects with many boundary edges such as laptops, tables, and chairs.
On S3DIS Area 5, BAGNet reaches 56.5 mIoU and 80.2 ms inference time, trimming inference time by 13.2% to 77.9% relative to the compared baselines.
The ablation attributes roughly 3.9% mIoU to the BAGLayer, about 2.2% to the attention pooling layer, and 6.1% to the combination, with extra latency of only 4.4 ms and 1.9 ms respectively.
Concentrating graph computation on boundary points lets the network keep graph-based expressiveness where it matters while using a cheap MLP elsewhere, which is the reported mechanism behind the speed-accuracy trade-off.
The attention pooling layer restores global context at low cost, and the paper shows it improves boundary-point mIoU as well as overall mIoU.

Reading between the lines

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

If the boundary detector remains robust across densities and scene types, the same split-and-specialize recipe could transfer to instance segmentation or object detection, where edge regions are also the hardest cases.
The paper reports overall accuracy on S3DIS of 84.5%, slightly below GAPointNet's 85.0%, so the practical advantage may rest specifically on mIoU and latency rather than on every accuracy metric.
Making the boundary detector's threshold explicit and testing sensitivity over threshold values would clarify how much of the measured gain depends on detector calibration.
One could measure the fraction of points labeled boundary per object class and correlate that fraction with per-class mIoU gains, directly testing the explanation that edge-vertex fusion helps boundary-rich objects most.

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.

Desk Editor's Note

read the letter

My quick take: BAGNet’s central idea—spend graph attention only on boundary points and use a cheap MLP elsewhere—is sensible and the ablation suggests it works. The paper’s problem is not the idea but the missing specification of what makes a point a boundary point. Section III-A says only 'boundary estimation based on normal vectors' and gives no algorithm, threshold, or validation. Since the accuracy gain and the latency win both rest on that partition, the reported 86.2 mIoU on ShapeNet and 80.2 ms on S3DIS are not reproducible as written. That’s a load-bearing gap, not a stylistic complaint.

The comparisons also omit the recent transformer baselines the paper itself cites (Point Transformer, Stratified Transformer, Fast Point Transformer), so the 'state-of-the-art' label is overclaimed. No error bars, no repeated seeds, and the S3DIS table has only a handful of methods. The attention pooling layer is described as producing a (k,512) global vector from (N,k,512) features; that shape only makes sense if it’s max over N, which isn’t a global descriptor. Eq. (4) indexes y'_i by j up to k while Eq. (3) defines y'_i as a D-vector; presumably a k-dim MLP is intended but not written. These are fixable in revision.

On the credit side: the paper does give an ablation showing BAGLayer and pooling each add a few mIoU points, with modest inference time cost, and the latency numbers in Table II are at least internally consistent with the design. The self-citations [2]–[4] are background, not a problem.

I’d send this to review with a clear request for major revision: specify the boundary detector, add modern baselines, fix the notation, and release code or detailed hyperparameters. Without those, the claims are unverifiable. I wouldn’t cite it in its current form, but the idea is worth watching.

Referee Report

5 major / 5 minor

Summary. The paper introduces BAGNet, a boundary-aware graph attention network for 3D point cloud semantic segmentation. It partitions points into boundary and non-boundary points using an unspecified normal-vector-based boundary estimation, applies a boundary-aware graph attention layer (BAGLayer) to boundary points, a pointwise MLP to non-boundary points, and an attention pooling layer to the raw point cloud to obtain a global feature; these features are concatenated and classified by a shared MLP. Experiments on ShapeNet Part and S3DIS Area 5 report 86.2 and 56.5 mIoU, respectively, and an S3DIS inference time of 80.2 ms, both better than the listed baselines. An ablation study reports that the BAGLayer and the attention pooling layer each contribute to accuracy. The paper does not release code and leaves key implementation details, especially the boundary detector, unspecified.

Significance. If the claims hold, BAGNet offers a practically attractive trade-off: it concentrates graph attention on boundary points and uses cheap processing elsewhere, which could explain the simultaneous accuracy and speed improvements. The paper's strengths are the clear geometric motivation, the modular architecture, the standard benchmark evaluation, and the inclusion of ablations for both modules. However, the central accuracy and latency claims are currently not reproducible because the boundary partition is described only qualitatively, and the technical exposition contains dimensional and algebraic inconsistencies in the attention pooling and BAGLayer formulations. There are no machine-checked proofs or released code to compensate. With the required specification and corrections, the core idea is worth publishing; in its current form, the missing boundary-detection protocol is a load-bearing gap.

major comments (5)

[Section III-A] The entire architecture and both headline claims depend on the partition of points into boundary and non-boundary points, but Section III-A only states that this is done 'according to boundary estimation based on normal vectors' and lists two qualitative criteria. No algorithm, threshold, detector neighborhood, or validation against semantic labels is given. Figure 1 is illustrative only. Consequently, the 86.2 mIoU on ShapeNet and the 80.2 ms inference time in Tables I and II cannot be reproduced or attributed to BAGNet as specified. Please provide the exact boundary-detection procedure, report its precision/recall (or an ablation over the detection threshold), and state how the detected boundary fraction affects runtime.
[Section III-C] The attention pooling layer description is dimensionally inconsistent. The text states that KNN changes the input from (N,3) to (N,k,3), the MLP produces (N,k,512), and then 'a max pooling operation is applied ... resulting in a global feature vector with a shape of (k,512)'. If the max is taken over the N points, the result should be a (1,512) or (512,) global vector; a (k,512) shape is not global. If the max is taken over the neighbor dimension instead, it does not aggregate all points. The layer is also called an attention pooling layer, but only max pooling is described. Please correct the shape, specify the pooling axis, and either define the attention weights or rename the layer.
[Section III-B, Eqs. (3)-(4)] Equation (3) defines y'_i = h(hat p'_i + sum_j hat p'_ij - sum_j bar W'_ij), so y'_i is a D-dimensional vector (D=256), but Equation (4) treats y'_ij as the j-th component of y'_i for j=1,...,k with k=32. This is dimensionally inconsistent unless D=k, which contradicts the stated configuration. In addition, the paper says the subtraction of the edge terms removes the double-counted edge-weight contribution, but the edge weights were added before the three MLP encodings; subtracting the encoded edge features after the nonlinearity does not cancel the earlier additive contributions. Please rewrite Equation (3) to produce per-neighbor attention logits and either derive the claimed cancelation under explicit linear encodings or remove the claim.
[Section IV-A] The implementation details say 'For feature extraction of non-boundary points, we utilize PointNet++ [6] in the pointwise MLP', but PointNet++ is a hierarchical network, not a pointwise MLP, and the rest of the paper describes non-boundary points as being processed by a simple shared MLP. This is load-bearing because the claimed speed advantage relies on the non-boundary path being cheap. Please specify the exact operator used for non-boundary points and explain how its features are combined with the BAGLayer output.
[Section IV-B, Table II] The paper's abstract claims 'less inference time', and Table II reports 80.2 ms for BAGNet, but no measurement protocol is given: no hardware, software framework, batch size, or statement of whether the reported time includes boundary estimation, KNN construction, and preprocessing. Without this information, the reported 13.2% to 77.9% reduction in inference time cannot be verified or compared fairly with the baselines. Please report the full measurement setup and, ideally, a breakdown of where time is spent.

minor comments (5)

[Table I] The SGPN row contains corrupted entries such as '95.877.893.8' and '60.192.389.4'; the entire table should be reformatted for readability.
[Figure 5] The configurations of Models A-D are presented only inside the figure; the text should state which modules each model contains and how the boundary IoU metric is defined.
[Section III-B] The caption of Figure 3 refers to a 'single channel BAGLayer', but no multi-channel variant is described in the text; please define what a channel is or remove the term.
[Section IV-B] The heading 'Comparsion Experiment Results' contains a typo and should read 'Comparison Experiment Results'.
[Section III-C] The paragraph describing the KNN network says it 'extracts the local geometric features of each point', but the described operation only rearranges coordinates into (N,k,3); the feature extraction happens in the subsequent MLP. Please rephrase to avoid conflating these steps.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: BAGNet's accuracy and speed results are evaluated against external benchmarks and baselines, and the boundary-point routing is an upstream design assumption rather than a fitted quantity renamed as a prediction.

full rationale

The paper's central claim is an empirical performance comparison on public datasets (ShapeNet and S3DIS), with mIoU and inference time reported against independent baseline methods. No equation in the method section is fit to those reported results, and no reported number is derived by construction from a parameter fitted on the test set. The boundary/non-boundary partition in Section III-A is described only as 'boundary estimation based on normal vectors,' which is a missing-implementation and missing-validation concern, not a circularity concern: the detector is not defined in terms of the semantic labels it is used to improve, and the reported accuracy is measured against external ground truth. The ablation study compares models with and without the proposed modules on the same dataset; such an ablation can be unpersuasive if hyperparameters or evaluation protocols are cherry-picked, but it does not reduce the central claim to its inputs. The self-citations [2]-[4] appear in the introduction as examples of point cloud applications and are not load-bearing for BAGNet's architecture or its claimed superiority. No uniqueness theorem, ansatz, or fitted constant is imported from the authors' prior work. Therefore, while the paper has serious reproducibility weaknesses, none of them constitutes circular reasoning under the specified criteria.

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

All entries are hyperparameters or unstated modeling choices. The ledger shows the method is an architecture with several hand-chosen constants rather than a derivation with fitted physical quantities.

free parameters (4)

Boundary detection threshold = not reported
The partition into boundary and non-boundary points is decided by an unspecified normal-vector criterion; this is a hand-chosen model of what counts as a boundary and materially affects where graph attention is spent.
Neighbor count k = 32
Used for KNN graph construction in both BAGLayer and attention pooling; chosen by hand.
BAGLayer output dimension D = 256
Feature dimension of the boundary feature representation, chosen by hand.
Attention pooling MLP widths = (16,64,128,512)
Architecture of the global feature extractor, chosen by hand.

assumptions (4)

domain assumption Boundary points exhibit chaotic normal-vector distributions and carry more structural information than non-boundary points.
Central motivation in Section I and Figure 1; the paper does not provide quantitative evidence or a segmentation-boundary benchmark for this claim.
domain assumption Non-boundary points are simple enough for a pointwise MLP to segment correctly.
The speed gain assumes cheap MLP features suffice for non-boundary regions (Section III-A).
ad hoc to paper Subtracting the encoded edge features in Eq (3) removes the double-counted edge weight contribution.
The subtraction happens after nonlinear MLP encodings, so it does not linearly cancel the duplicated edge weight; the paper asserts it reduces the impact without proof (Section III-B, Eq. 3).
standard math KNN graphs capture the local connectivity needed for semantic segmentation.
Standard practice; used to build the graph in Section III-B.

how reviews work

0 comments

Cite this review

Pith. "Pith review of BAGNet: A Boundary-Aware Graph Attention Network for 3D Point Cloud Semantic Segmentation." pith.science (2026). https://pith.science/paper/3IE2SMCK

@misc{pith2026250600475,
  author       = {Pith},
  title        = {Pith review of: BAGNet: A Boundary-Aware Graph Attention Network for 3D Point Cloud Semantic Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3IE2SMCK}},
  note         = {Machine review of arXiv:2506.00475}
}

read the original abstract

Since the point cloud data is inherently irregular and unstructured, point cloud semantic segmentation has always been a challenging task. The graph-based method attempts to model the irregular point cloud by representing it as a graph; however, this approach incurs substantial computational cost due to the necessity of constructing a graph for every point within a large-scale point cloud. In this paper, we observe that boundary points possess more intricate spatial structural information and develop a novel graph attention network known as the Boundary-Aware Graph attention Network (BAGNet). On one hand, BAGNet contains a boundary-aware graph attention layer (BAGLayer), which employs edge vertex fusion and attention coefficients to capture features of boundary points, reducing the computation time. On the other hand, BAGNet employs a lightweight attention pooling layer to extract the global feature of the point cloud to maintain model accuracy. Extensive experiments on standard datasets demonstrate that BAGNet outperforms state-of-the-art methods in point cloud semantic segmentation with higher accuracy and less inference time.

Figures

Figures reproduced from arXiv: 2506.00475 by the authors.

**Figure 2.** The overall architecture of our network: We divide the entire point [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** The architecture of the single channel BAGLayer: We select a boundary point as the central point and use the KNN algorithm to select [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: The architecture of the attention pooling layer. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: The semantic segmentation results of different models on the dataset [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 30 canonical work pages

[2]

Pointactionclip: Preventing transfer degradation in point cloud action recognition with a triple-path clip,

W. Tao, S. He, X. Qu, J. Wan, and J. Wang, “Pointactionclip: Preventing transfer degradation in point cloud action recognition with a triple-path clip,” in2025 IEEE International Conference on Acoustics, Speech, and Signal Processing, 2025

work page 2025
[4]

3d point cloud segmentation for complex structure based on pointsift,

Z. Li, J. Wang, X. Qu, and J. Xiao, “3d point cloud segmentation for complex structure based on pointsift,” inChinese Conference on Pattern Recognition and Computer Vision (PRCV), pp. 660–670, Springer, 2020

work page 2020
[6]

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

C. R. Qi, L. Yi, H. Su, and L. J. Guibas, “Pointnet++: Deep hierarchical feature learning on point sets in a metric space,”Advances in Neural Information Processing Systems, vol. 30, 2017

2017
[1]

Point cloud denoising using normal vector-based graph wavelet shrinkage,

R. Watanabe, K. Nonaka, H. Kato, E. Pavez, T. Kobayashi, and A. Or- tega, “Point cloud denoising using normal vector-based graph wavelet shrinkage,” inICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 2569–2573, IEEE, 2022

work page 2022
[3]

Prenet: A plane- fit redundancy encoding point cloud sequence network for real-time 3d action recognition,

S. He, X. Qu, J. Wan, G. Li, C. Xie, and J. Wang, “Prenet: A plane- fit redundancy encoding point cloud sequence network for real-time 3d action recognition,” in2024 International Joint Conference on Neural Networks (IJCNN), pp. 1–8, IEEE, 2024

work page 2024
[5]

Pointnet: Deep learning on point sets for 3d classification and segmentation,

C. R. Qi, H. Su, K. Mo, and L. J. Guibas, “Pointnet: Deep learning on point sets for 3d classification and segmentation,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 652–660, 2017

work page 2017
[7]

Gapointnet: Graph attention based point neural network for exploiting local feature of point cloud,

C. Chen, L. Z. Fragonara, and A. Tsourdos, “Gapointnet: Graph attention based point neural network for exploiting local feature of point cloud,” Neurocomputing, vol. 438, pp. 122–132, 2021

work page 2021
[8]

Dynamic graph cnn for learning on point clouds,

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, vol. 38, no. 5, pp. 1–12, 2019

work page 2019

Show all 33 references

[9]

Spherical kernel for efficient graph convolution on 3d point clouds,

H. Lei, N. Akhtar, and A. Mian, “Spherical kernel for efficient graph convolution on 3d point clouds,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 43, no. 10, pp. 3664–3680, 2020

2020
[10]

Adaptive graph convolution for point cloud analysis,

H. Zhou, Y . Feng, M. Fang, M. Wei, J. Qin, and T. Lu, “Adaptive graph convolution for point cloud analysis,” inProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 4965–4974, 2021

2021
[11]

Improving graph representation for point cloud segmentation via attentive filtering,

N. Zhang, Z. Pan, T. H. Li, W. Gao, and G. Li, “Improving graph representation for point cloud segmentation via attentive filtering,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 1244–1254, 2023

2023
[12]

Exploring spatial context for 3d semantic segmentation of point clouds,

F. Engelmann, T. Kontogianni, A. Hermans, and B. Leibe, “Exploring spatial context for 3d semantic segmentation of point clouds,” in Proceedings of the IEEE International Conference on Computer Vision Workshops, pp. 716–724, 2017

2017
[13]

3d recurrent neural networks with context fusion for point cloud semantic segmentation,

X. Ye, J. Li, H. Huang, L. Du, and X. Zhang, “3d recurrent neural networks with context fusion for point cloud semantic segmentation,” in Proceedings of the European Conference on Computer Vision (ECCV), pp. 403–417, 2018

2018
[14]

Randla-net: Efficient semantic segmentation of large- scale point clouds,

Q. Hu, B. Yang, L. Xie, S. Rosa, Y . Guo, Z. Wang, N. Trigoni, and A. Markham, “Randla-net: Efficient semantic segmentation of large- scale point clouds,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11108–11117, 2020

2020
[15]

Point convolutional neural networks by extension operators,

M. Atzmon, H. Maron, and Y . Lipman, “Point convolutional neural networks by extension operators,”ACM Transactions on Graphics, vol. 37, no. 4, pp. 1–12, 2018

2018
[16]

Pointwise convolutional neural networks,

B.-S. Hua, M.-K. Tran, and S.-K. Yeung, “Pointwise convolutional neural networks,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 984–993, 2018

2018
[17]

A-cnn: Annularly convolutional neural networks on point clouds,

A. Komarichev, Z. Zhong, and J. Hua, “A-cnn: Annularly convolutional neural networks on point clouds,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 7421– 7430, 2019

2019
[18]

Pointcnn: Convolution on x-transformed points,

Y . Li, R. Bu, M. Sun, W. Wu, X. Di, and B. Chen, “Pointcnn: Convolution on x-transformed points,”Advances in Neural Information Processing Systems, vol. 31, 2018

2018
[19]

Semi-supervised classification with graph convolutional networks,

T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” inInternational Conference on Learning Rep- resentations, 2017

2017
[20]

Attention is all you need,

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, vol. 30, 2017

2017
[21]

Stratified transformer for 3d point cloud segmentation,

X. Lai, J. Liu, L. Jiang, L. Wang, H. Zhao, S. Liu, X. Qi, and J. Jia, “Stratified transformer for 3d point cloud segmentation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 8500–8509, 2022

2022
[22]

Pyramid architecture for multi- scale processing in point cloud segmentation,

D. Nie, R. Lan, L. Wang, and X. Ren, “Pyramid architecture for multi- scale processing in point cloud segmentation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 17284–17294, 2022

2022
[23]

Fast point transformer,

C. Park, Y . Jeong, M. Cho, and J. Park, “Fast point transformer,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 16949–16958, 2022

2022
[24]

Learning inner-group relations on point clouds,

H. Ran, W. Zhuo, J. Liu, and L. Lu, “Learning inner-group relations on point clouds,” inProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 15477–15487, 2021

2021
[25]

Patchformer: An efficient point transformer with patch attention,

C. Zhang, H. Wan, X. Shen, and Z. Wu, “Patchformer: An efficient point transformer with patch attention,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11799– 11808, 2022

2022
[26]

Point transformer,

H. Zhao, L. Jiang, J. Jia, P. H. Torr, and V . Koltun, “Point transformer,” inProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 16259–16268, 2021

2021
[27]

Less is more: label recommendation for weakly supervised point cloud semantic segmenta- tion,

Z. Pan, N. Zhang, W. Gao, S. Liu, and G. Li, “Less is more: label recommendation for weakly supervised point cloud semantic segmenta- tion,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, pp. 4397–4405, 2024

2024
[28]

Escape from cells: Deep kd-networks for the recognition of 3d point cloud models,

R. Klokov and V . Lempitsky, “Escape from cells: Deep kd-networks for the recognition of 3d point cloud models,” inProceedings of the IEEE International Conference on Computer Vision, pp. 863–872, 2017

2017
[29]

Recurrent slice networks for 3d segmentation of point clouds,

Q. Huang, W. Wang, and U. Neumann, “Recurrent slice networks for 3d segmentation of point clouds,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2626–2635, 2018

2018
[30]

Sgpn: Similarity group proposal network for 3d point cloud instance segmentation,

W. Wang, R. Yu, Q. Huang, and U. Neumann, “Sgpn: Similarity group proposal network for 3d point cloud instance segmentation,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2569–2578, 2018

2018
[31]

Shapenet: An information- rich 3d model repository,

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
[32]

3d semantic parsing of large-scale indoor spaces,

I. Armeni, O. Sener, A. R. Zamir, H. Jiang, I. Brilakis, M. Fischer, and S. Savarese, “3d semantic parsing of large-scale indoor spaces,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 1534–1543, 2016

2016
[33]

Segcloud: Semantic segmentation of 3d point clouds,

L. Tchapmi, C. Choy, I. Armeni, J. Gwak, and S. Savarese, “Segcloud: Semantic segmentation of 3d point clouds,” in2017 International Conference on 3D Vision (3DV), pp. 537–547, IEEE, 2017

2017

Pith tools

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

[1] [2]

Pointactionclip: Preventing transfer degradation in point cloud action recognition with a triple-path clip,

W. Tao, S. He, X. Qu, J. Wan, and J. Wang, “Pointactionclip: Preventing transfer degradation in point cloud action recognition with a triple-path clip,” in2025 IEEE International Conference on Acoustics, Speech, and Signal Processing, 2025

work page 2025

[2] [4]

3d point cloud segmentation for complex structure based on pointsift,

Z. Li, J. Wang, X. Qu, and J. Xiao, “3d point cloud segmentation for complex structure based on pointsift,” inChinese Conference on Pattern Recognition and Computer Vision (PRCV), pp. 660–670, Springer, 2020

work page 2020

[3] [6]

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

C. R. Qi, L. Yi, H. Su, and L. J. Guibas, “Pointnet++: Deep hierarchical feature learning on point sets in a metric space,”Advances in Neural Information Processing Systems, vol. 30, 2017

2017

[4] [1]

Point cloud denoising using normal vector-based graph wavelet shrinkage,

R. Watanabe, K. Nonaka, H. Kato, E. Pavez, T. Kobayashi, and A. Or- tega, “Point cloud denoising using normal vector-based graph wavelet shrinkage,” inICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 2569–2573, IEEE, 2022

work page 2022

[5] [3]

Prenet: A plane- fit redundancy encoding point cloud sequence network for real-time 3d action recognition,

S. He, X. Qu, J. Wan, G. Li, C. Xie, and J. Wang, “Prenet: A plane- fit redundancy encoding point cloud sequence network for real-time 3d action recognition,” in2024 International Joint Conference on Neural Networks (IJCNN), pp. 1–8, IEEE, 2024

work page 2024

[6] [5]

Pointnet: Deep learning on point sets for 3d classification and segmentation,

C. R. Qi, H. Su, K. Mo, and L. J. Guibas, “Pointnet: Deep learning on point sets for 3d classification and segmentation,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 652–660, 2017

work page 2017

[7] [7]

Gapointnet: Graph attention based point neural network for exploiting local feature of point cloud,

C. Chen, L. Z. Fragonara, and A. Tsourdos, “Gapointnet: Graph attention based point neural network for exploiting local feature of point cloud,” Neurocomputing, vol. 438, pp. 122–132, 2021

work page 2021

[8] [8]

Dynamic graph cnn for learning on point clouds,

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, vol. 38, no. 5, pp. 1–12, 2019

work page 2019