Pith. sign in

REVIEW 4 major objections 5 minor 27 references

XAI for Point Cloud Data using Perturbations based on Meaningful Segmentation

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Explaining point-cloud classifiers with semantic part segments yields saliency maps humans can read.

desk verdict A genuinely new segmentation-based perturbation method for point cloud XAI that works for PointNet, but the model-agnostic claim is unsupported because the point-shifting invariance is architecture-specific. read the letter →

arxiv 2507.22020 v1 pith:LA4OJJRF submitted 2025-07-29 cs.CV cs.AI

classification cs.CVcs.AI
keywords explainableAIpointcloudclassificationsaliencymapssemanticsegmentationperturbation-basedXAIshiftingmodel-agnosticexplanationpart
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that explanations of point cloud classifiers are only as good as the units used to perturb the input. It proposes building saliency maps from meaningful segments produced by part-segmentation models, rather than from arbitrary clusters, and introduces a point-shifting perturbation that moves a segment's points to a random point inside the retained structure so the shifted points no longer contribute structural information. The two resulting saliency modes, removing one segment or keeping one segment while removing the rest, give complementary views of what a classifier uses. A sympathetic reader would care because the method is model-agnostic and produces attributions that a human can attach to named object parts, which clustering-based approaches cannot do.

What carries the argument

The load-bearing mechanism is the point-shifting perturbation: instead of moving a segment's points to the cloud center, which can create a new dense artifact, the points are shifted to a random point belonging to the retained structure so the relocated points merge into an existing part and contribute no additional structure. The second building block is semantic segmentation by trained per-class part-segmentation models, which defines the perturbation regions. The saliency attributions are computed by the equations $S_{AF}(x) = |P(a) - P(a')|$ and $S_{PF}(x) = -|P(a) - P(a'')|$, where $P$ is the classifier's target-class score, $a$ the original input, and $a'$/$a''$ the perturbed inputs.

What would settle it

Take a fixed input cloud, choose a segment, and shift its points to each of many random points in the retained structure; if the classifier's full output vector varies with the chosen target point for even a small fraction of segments, then the shifted points still carry structural influence and the computed attributions are contaminated. A quantitative version would report the distribution of output changes across, say, 100 random target points per segment and per object.

Watch

Extended reading notes

Core claim

The central claim is that segmentation-based perturbation produces more meaningful, interpretable saliency maps for point cloud classification than clustering-based perturbation. Meaningful segments are obtained by training per-category part-segmentation models on the same point clouds, so each segment corresponds to a recognizable object part such as a wing or a wheel. Perturbations are then applied through a point-shifting mechanism that relocates all points of a segment to a randomly chosen point that is itself part of the retained structure. The paper claims this placement makes the shifted points inert: they add no new structural information and leave the classifier's output vector unchanged. Saliency attributions are computed from the change in target-class score under two schemes, absence of a feature where a single segment is shifted and presence of a feature where all other segments are shifted, and the paper demonstrates on examples that the resulting maps separate meaningful parts and even reveal per-part differences such as front versus rear wheel influence on a motorbike classification.

Load-bearing premise

The method hinges on the assumption that shifting a segment's points to a random point inside the retained structure leaves the classifier's output completely unchanged, and this is checked only visually on two airplane examples without statistical verification.

Editorial extensions

If this is right

  • Saliency maps produced this way attach attributions to named object parts, so a user can read an explanation as 'the wings drove this prediction' rather than 'cluster 3 drove this prediction'.
  • The two perturbation modes offer complementary views: absence shows what the model misses when a part disappears, while presence shows how much a part alone supports the predicted class.
  • Because the approach is model-agnostic at the explanation stage, it can be applied to any point cloud classifier, not just the architecture used in the experiments.
  • Further splitting segments with clustering yields finer attributions, e.g., separating the front from the rear wheel of a motorbike, which a pure segmentation cannot do.
  • The method tolerates moderate input noise, up to roughly 10 percent, before segmentation errors degrade the saliency maps.

Reading between the lines

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

  • If the random-point inertness assumption holds, the same perturbation trick could be reused in other XAI schemes, such as Shapley-value estimation over semantic parts, to make those explanations interpretable too.
  • The dependence on per-category segmentation models is the main scaling bottleneck; a single open-vocabulary or zero-shot part segmenter would remove the need for labeled part data when adding new object categories.
  • The two perturbation modes resemble counterfactual reasoning: presence-of-feature is close to asking 'would this part alone still be classified as this object?', and could be quantified as a minimal-support measure for the class.
  • A statistical stability test of attributions across random target points would turn the current anecdotal verification into a reproducible guarantee and could serve as a quality metric for perturbation-based point-cloud XAI.
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 a perturbation-based XAI method for point cloud classification networks, in which semantic segmentation models define perturbation regions (segments) and saliency is computed from the change in the target class score when a segment is removed (Equation 1) or when all but a segment are removed (Equation 2). A point-shifting mechanism shifts segment points onto a randomly chosen point of the retained structure, with the aim of making the shifted points structurally inert. The method is evaluated with qualitative examples on PointNet classifiers and per-category PointNet-based segmentation models, and is compared qualitatively with k-means clustering. The central claim is that segmentation-based perturbations produce more meaningful, human-interpretable saliency maps than clustering-based methods, and that the method is model-agnostic and architecture-independent.

Significance. If the central claim holds, the paper addresses a real gap: most point cloud XAI methods use geometric clusters or individual points, whose semantic meaning is often unclear to users. Using semantic part segments as perturbation units is a reasonable and potentially impactful idea, and the point-shifting mechanism is an interesting attempt to avoid the artifacts of deleting or clustering points. The paper also makes a useful effort to compare segmentation-based perturbations with clustering-based ones and to study the effect of segmentation accuracy and input noise. However, the current evidence is almost entirely qualitative, and the load-bearing no-influence assumption of the point-shifting mechanism is validated only anecdotally. The significance of the contribution therefore depends on whether the invariance claim can be established for the architectures the authors address and whether 'meaningfulness' can be demonstrated beyond subjective visual inspection.

major comments (4)
  1. [§3.2.2 and §3.4] The load-bearing claim that shifting segment points to a random point of the retained structure makes them 'no longer influence the output' is not established. Section 4.2 reports that for two airplane examples 'the output values (all 16 values in the output vector) did not change', but no actual output vectors, numeric values, or statistical evidence are provided. More importantly, the stated property is architecture-dependent: it holds for PointNet because per-point MLPs followed by global max-pooling are invariant to duplicate coordinates, but for models with local aggregation or neighborhood operations (for example PointNet++, point transformers, or graph networks), duplicate points change local density, grouping, and farthest-point sampling. The paper's claim in Section 5 that the method is model-agnostic 'irrespective of the model architecture' is therefore unsupported, and for such architectures the saliency attributions computed by Equations (1) and (2) would mix the intended segment-deletion effect with spurious artifacts from duplicated points. The authors should either validate the invariance empirically across a range of architectures or substantially restrict the model-agnostic claim.
  2. [§3.2.2/§3.4] There is an internal inconsistency between the two perturbation mechanisms. Section 3.4 argues that shifting points to the center of the point cloud is undesirable because the center may act as an additional artificial structure, and Figure 7 illustrates this failure mode. However, the 'presence of a feature' mechanism described in Section 3.2.2 and shown in Figure 5 moves all non-retained points to the center of the point cloud. If the argument of Section 3.4 is correct, then the perturbations used in Equation (2) introduce exactly the kind of spurious structure that the point-shifting mechanism was designed to avoid. The paper does not explain why this objection does not apply to the 'presence of a feature' variant, or whether the random-point mechanism is in fact used there. This needs clarification, and if the center-shifting is indeed used, the attributions from Equation (2) are contaminated by an additional structural feature.
  3. [§4.1-§4.5] The central claim of generating 'meaningful', easily interpretable saliency maps is supported only by qualitative visual inspection of a small number of examples. There is no quantitative evaluation of faithfulness, stability, or human interpretability, and no user study. The comparison with clustering-based methods in Section 4.1 is also qualitative: the observation that saliency maps vary with the number of clusters is made for a few examples without quantitative measures of variance or reliability. To support the paper's main contribution, the authors should provide quantitative evidence, such as insertion/deletion or fidelity metrics against the classifier, stability measures across random choices of the shifting point, and ideally a user study comparing semantic and clustering-based saliency maps.
  4. [§5] The statement that the proposed method is model-agnostic because it is perturbation-based is too strong given the experimental scope. All experiments use PointNet-based classification and segmentation models, and the point-shifting mechanism's invariance relies on PointNet's global max-pooling structure, as noted above. The paper does not demonstrate the method on any other architecture, so the 'irrespective of the model architecture' claim in Section 5 is not supported by the evidence. The authors should either add experiments on at least one non-max-pooling architecture or revise the claim to state the method is perturbation-based and can in principle be applied to any model, subject to the validity of the invariance assumption.
minor comments (5)
  1. [§3.2.2] The use of the minus sign in Equation (2) and the statement that it 'allows the segment having the highest influence on the output value to have the highest attribution' is unclear: with the formula as written, a segment that causes a large output change receives a more negative value, so the mapping from value to visualization should be specified explicitly.
  2. [§4.2] The observation that 'the output values did not change' should be substantiated with a table of the actual output vectors or a statistical summary over a set of random points and multiple inputs, rather than a single anecdotal statement about two perturbations of one airplane.
  3. [§4.5.2] The claim that the method produces 'similar saliency attributions' for inputs with up to 10% noise is not quantified; reporting a numeric similarity measure, such as mean absolute difference or rank correlation between saliency vectors, would make this statement precise.
  4. [§3.1.1] The selection of the segmentation model based on the classifier's predicted class is a dependency that the paper acknowledges in Section 4.6, but it should also be mentioned in Section 3 as a potential source of error in the overall pipeline, since a misclassification will select an inappropriate segmentation model.
  5. [§4.3] The discussion of the wing/fuselage saliency differences across airplane examples is interesting but would benefit from a quantitative validation, e.g., reporting the actual saliency values for wings and fuselage for the eight examples in Figure 10, rather than relying only on color scales.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the saliency values are defined directly from classifier output changes over independently trained semantic segments, with no fitted parameters or self-referential derivation chain.

full rationale

The paper's saliency computation is a standard perturbation-based attribution: Equation (1) and Equation (2) define attributions as the change in the classifier's output score after perturbing a segment, which is exactly the definition of the method rather than a hidden reuse of its inputs. The segments are produced by pre-trained PointNet segmentation models trained on ShapeNet part labels, i.e., an externally supplied semantic decomposition, not derived from the saliency values or from the explained classifier. The 'meaningfulness' of the resulting maps is inherited from the supervised segmentation input, but this is an explicit design choice and not a tautological derivation: the attribution magnitudes still depend on the classifier's response to each perturbed input. No parameter is fitted to the saliency maps, and no 'prediction' is made from a fitted subset of the same data. The only self-citation, the authors' own survey [11], appears in the related-work discussion and is not load-bearing for any technical claim. The point-shifting assumption that shifted points no longer influence the output is empirically under-supported and may fail for architectures with local aggregation, but that is a correctness and generalization limitation, not a circularity. The paper is otherwise self-contained against external models and data, so the appropriate finding is no significant circularity.

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

The method introduces no fitted parameters or new entities. It relies on domain assumptions about the point-shifting mechanism, the meaningfulness of semantic segments, and the correctness of the classifier for model selection.

assumptions (3)
  • domain assumption Shifting points of a segment to a random point in the retained structure removes their influence on the classifier output.
    Section 3.4 states this without quantitative proof; the entire perturbation mechanism relies on it.
  • domain assumption Segments produced by the pretrained segmentation models are meaningful to humans.
    Section 3.1 assumes part labels correspond to intuitive object parts (e.g., head, hands, legs, torso); no user validation is provided.
  • domain assumption The classification model's predicted class is correct enough to select the appropriate segmentation model.
    Section 3 describes selecting the segmentation model based on the output class; Section 4.6 notes this fails when misclassification occurs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of XAI for Point Cloud Data using Perturbations based on Meaningful Segmentation." pith.science (2026). https://pith.science/paper/LA4OJJRF

@misc{pith2026250722020,
  author       = {Pith},
  title        = {Pith review of: XAI for Point Cloud Data using Perturbations based on Meaningful Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LA4OJJRF}},
  note         = {Machine review of arXiv:2507.22020}
}
read the original abstract

We propose a novel segmentation-based explainable artificial intelligence (XAI) method for neural networks working on point cloud classification. As one building block of this method, we propose a novel point-shifting mechanism to introduce perturbations in point cloud data. Recently, AI has seen an exponential growth. Hence, it is important to understand the decision-making process of AI algorithms when they are applied in critical areas. Our work focuses on explaining AI algorithms that classify point cloud data. An important aspect of the methods used for explaining AI algorithms is their ability to produce explanations that are easy for humans to understand. This allows them to analyze the AI algorithms better and make appropriate decisions based on that analysis. Therefore, in this work, we intend to generate meaningful explanations that can be easily interpreted by humans. The point cloud data we consider represents 3D objects such as cars, guitars, and laptops. We make use of point cloud segmentation models to generate explanations for the working of classification models. The segments are used to introduce perturbations into the input point cloud data and generate saliency maps. The perturbations are introduced using the novel point-shifting mechanism proposed in this work which ensures that the shifted points no longer influence the output of the classification algorithm. In contrast to previous methods, the segments used by our method are meaningful, i.e. humans can easily interpret the meaning of the segments. Thus, the benefit of our method over other methods is its ability to produce more meaningful saliency maps. We compare our method with the use of classical clustering algorithms to generate explanations. We also analyze the saliency maps generated for example inputs using our method to demonstrate the usefulness of the method in generating meaningful explanations.

Figures

Figures reproduced from arXiv: 2507.22020 by the authors.

Figure 1
Figure 1. An overview of the XAI pipeline proposed. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 1
Figure 1. It consists of four stages: • Classification • Segmentation • Perturbation • Saliency mapping In the first stage, the input point cloud data is used as the input for the classification model which predicts the output class of this data. This is the same classification model that we intend to understand in the XAI process. Based on the output class, the corresponding segmentation model is chosen from the list of pre-… view at source ↗
Figure 2
Figure 2. Examples of point clouds segmentation performed by our segmentation models and their corresponding [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figures from the paper (12 more)
Figure 3
Figure 3. Figure 3: Clustering of the segments produced by the segmentation models to obtain finer segments. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png]
Figure 4
Figure 4. Figure 4: Saliency map produced by our method for the given input point cloud data representing a plane. Note: Refer [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Structural information carried by a point cloud data (left figure) and its perturbed data where all the points [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Saliency map produced by using the "presence of feature" method for the input used in Figure 4. Note: Refer [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Shifting points to the centroid of the input point cloud data. Points representing the seat (in (b)) and seat & [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Saliency maps produced by the absence of feature mechanism for clusters produced by KMeans clustering method. c represents the number of clusters. 11 [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Perturbation of the input data using random points (marked in (c) and (d)) selected from the retained structure. [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: Saliency maps of point clouds representing airplanes using the [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: Saliency maps of point clouds representing motorbikes using the [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: Saliency maps produced with ((c) & (d)) and without ((e) & (f)) using the segmentation ground truth labels [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]
Figure 13
Figure 13. Figure 13: Noisy input with 5% noise and its corresponding saliency attributions produced by our proposed method.Note: [PITH_FULL_IMAGE:figures/full_fig_p015_13.png]
Figure 14
Figure 14. Figure 14: Distribution of the training data. used in this work have accuracy values in the range of 65%-80%. This is mainly due to two reasons. The first reason corresponds to the imbalance in the dataset. The dataset contains varying numbers of samples representing individual …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 17 canonical work pages

  1. [1]

    Explainable artificial intelligence for machine learning-based photogrammetric point cloud classification

    Muhammed Enes Atik, Zaide Duran, and Dursun Zafer Seker. Explainable artificial intelligence for machine learning-based photogrammetric point cloud classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 17:5834–5846, 2024. doi:10.1109/JSTARS.2024.3370159

  2. [2]

    Recent advancements in learning algorithms for point clouds: An updated overview

    Elena Camuffo, Daniele Mari, and Simone Milani. Recent advancements in learning algorithms for point clouds: An updated overview. Sensors, 22(4), 2022. doi:10.3390/s22041357

  3. [3]

    Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis 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, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chic...

  4. [4]

    A density-based algorithm for discovering clusters in large spatial databases with noise

    Martin Ester, Hans-Peter Kriegel, Jörg Sander, and Xiaowei Xu. A density-based algorithm for discovering clusters in large spatial databases with noise. In Proceedings of the Second International Conference on Knowledge Discovery and Data Mining, KDD’96, page 226–231. AAAI Press, 1996

  5. [5]

    Deep learning for 3D point clouds: A survey

    Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3D point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(12):4338–4364,

  6. [6]

    Toward Explainable Metrology 4.0: Utilizing Explainable AI to Predict the Pointwise Accuracy of Laser Scanning Devices in Industrial Manufacturing, pages 479–501

    Eleni Lavasa, Christos Chadoulos, Athanasios Siouras, Ainhoa Etxabarri Llana, Silvia Rodríguez Del Rey, Theodore Dalamagas, and Serafeim Moustakidis. Toward Explainable Metrology 4.0: Utilizing Explainable AI to Predict the Pointwise Accuracy of Laser Scanning Devices in Industrial Manufacturing, pages 479–501. Springer Nature Switzerland, Cham, 2024. doi...

  7. [7]

    A unified approach to interpreting model predictions

    Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In I. Guyon, U. V on Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30, pages 4765–4774, 2017. URL: https://proceedings.neurips. cc/paper/2017/file/8a20a8621978632d76c43dfd28b...

  8. [8]

    Some methods for classification and analysis of multivariate observations

    James MacQueen. Some methods for classification and analysis of multivariate observations. In Proceedings of the fifth Berkeley symposium on mathematical statistics and probability, volume 1, pages 281–297. Oakland, CA, USA, 1967

Show all 27 references
  1. [9]

    BubblEX: An explainable deep learning framework for point-cloud classification

    Francesca Matrone, Marina Paolanti, Andrea Felicetti, Massimo Martini, and Roberto Pierdicca. BubblEX: An explainable deep learning framework for point-cloud classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 15:6571–6587, 2022. d...

  2. [10]

    Interpretable geometric deep learning via learnable randomness injection, 2023

    Siqi Miao, Yunan Luo, Mia Liu, and Pan Li. Interpretable geometric deep learning via learnable randomness injection, 2023. arXiv:2210.16966, doi:10.48550/arXiv.2210.16966

  3. [11]

    Explainable artificial intelligence (xai) for methods working on point cloud data: A survey

    Raju Ningappa Mulawade, Christoph Garth, and Alexander Wiebel. Explainable artificial intelligence (xai) for methods working on point cloud data: A survey. IEEE Access, 12:146830–146851, 2024. doi:10.1109/ACCESS. 2024.3472872

  4. [12]

    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), July 2017. doi:10.1109/CVPR.2017.16

  5. [13]

    Why Should I Trust You?

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. "Why Should I Trust You?": Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’16, page 1135–1144, New York, NY , USA, ...

  6. [14]

    From 3d point-cloud data to explainable geometric deep learning: State-of-the-art and future challenges

    Anna Saranti, Bastian Pfeifer, Christoph Gollob, Karl Stampfer, and Andreas Holzinger. From 3d point-cloud data to explainable geometric deep learning: State-of-the-art and future challenges. WIREs Data Mining and Knowledge Discovery, 14(6):e1554, 2024. doi:10.1002/widm.1554

  7. [15]

    L. S. Shapley. A value for n-person games. pages 307–318, 1953. doi:doi:10.1515/9781400881970-018

  8. [16]

    Interpreting representation quality of DNNs for 3D point cloud processing

    Wen Shen, Qihan Ren, Dongrui Liu, and Quanshi Zhang. Interpreting representation quality of DNNs for 3D point cloud processing. In M. Ranzato, A. Beygelzimer, Y . Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34...

  9. [17]

    Interpretable rotation-equivariant quaternion neural networks for 3D point cloud processing

    Wen Shen, Zhihua Wei, Qihan Ren, Binbin Zhang, Shikun Huang, Jiaqi Fan, and Quanshi Zhang. Interpretable rotation-equivariant quaternion neural networks for 3D point cloud processing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(5):3290–3304, 2024. doi:10...

  10. [18]

    Axiomatic attribution for deep networks

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In International Conference on Machine Learning, pages 3319–3328. PMLR, 2017. doi:10.48550/arXiv.1703.01365

  11. [19]

    PointMask: Towards interpretable and bias-resilient point cloud processing, 2020

    Saeid Asgari Taghanaki, Kaveh Hassani, Pradeep Kumar Jayaraman, Amir Hosein Khasahmadi, and Tonya Custis. PointMask: Towards interpretable and bias-resilient point cloud processing, 2020. arXiv:2007.04525, doi:10.48550/arXiv.2007.04525

  12. [22]

    Flow AM: Generating point cloud global explanations by latent alignment

    Hanxiao Tan. Flow AM: Generating point cloud global explanations by latent alignment. 2024. arXiv: 2404.18760, doi:10.48550/arXiv.2404.18760

  13. [23]

    Surrogate model-based explainability methods for point cloud NNs

    Hanxiao Tan and Helena Kotthaus. Surrogate model-based explainability methods for point cloud NNs. In 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages 2927–2936, 2022. doi:10.1109/WACV51458.2022.00298

  14. [24]

    Explainability-aware one point attack for point cloud neural networks

    Hanxiao Tan and Helena Kotthaus. Explainability-aware one point attack for point cloud neural networks. In 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 4570–4579, 2023. doi:10.1109/WACV56688.2023.00456

  15. [25]

    F.M. Verburg. Exploring explainability and robustness of point cloud segmentation deep learning model by visualization, February 2022. URL: http://essay.utwente.nl/89440/

  16. [26]

    A scalable active framework for region annotation in 3d shape collections

    Li Yi, Vladimir G Kim, Duygu Ceylan, I Shen, Mengyan Yan, Hao Su, ARCewu Lu, Qixing Huang, Alla Sheffer, Leonidas Guibas, et al. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (TOG), 35(6):210, 2016

  17. [27]

    Min Zhang, Haoxuan You, Pranav Kadam, Shan Liu, and C.-C. Jay Kuo. PointHop: An explainable machine learning method for point cloud classification. IEEE Transactions on Multimedia , 22(7):1744–1755, 2020. doi:10.1109/TMM.2019.2963592

  18. [28]

    Pointcloud saliency maps

    Tianhang Zheng, Changyou Chen, Junsong Yuan, Bo Li, and Kui Ren. Pointcloud saliency maps. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. doi:10.48550/arXiv. 1812.01687. 18

  19. [2021]

    doi:10.1109/TPAMI.2020.3005434

Pith tools

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