Pith. sign in

REVIEW 4 major objections 4 minor 50 references

VGER: Voxel-Guided Global Event Ranking for Event Cloud Attribution

T0 review · 4 major / 4 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 VGER resolves the granularity dilemma in event attribution by transferring regional voxel evidence to individual events, and reports improved deletion-curve faithfulness on every dataset-backbone pair tested.

desk verdict VGER is a well-constructed new attribution method for event clouds, but the claimed nine-setting consistency is currently undermined by per-setting hyperparameter tuning and a thin baseline comparison. read the letter →

arxiv 2608.01470 v1 pith:P6ECPX6Z submitted 2026-08-02 cs.CV cs.AI

Youxin Jiang , Baoheng Fu , Hongwei Ren , Xiangqian Wu This is my paper · ORCID
classification cs.CVcs.AI
keywords eventcamerascloudattributionsaliencymapsvoxelperturbationdeletioncurvespoint-basednetworksinterpretabilityglobalranking
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

The paper introduces VGER, a training-free attribution method for point-based event-cloud networks. Its central claim is that one global ranking of individual events can simultaneously identify prediction-critical events at the top and low-impact, removable events at the bottom. VGER fuses per-event gradient sensitivity with task-aware voxel perturbation evidence, then refines the ordering with local structural rank and occupancy terms. On three event-camera benchmarks and three backbones, the paper reports that VGER improves both the high-tail and low-tail deletion AUCs over the point-level saliency baseline in all nine settings. The two-tail deletion protocol is the paper's proposed way to test attribution faithfulness from both ends of the same ordering.

What carries the argument

The load-bearing mechanism is the voxel-to-event evidence transfer: each occupied voxel is perturbed by contracting its events toward the robust stream center (preserving count, polarity, temporal order), and the resulting task-loss increase $\Delta L(V_j)$ is broadcast to every event in that voxel. This voxel channel supplies regional structure; an anisotropic center-contraction path $\bar{z}_i(\tau)=\exp(-\tau A_\rho)\bar{z}_i$, with $A_\rho=\mathrm{diag}(1,1,\rho)$, supplies per-event gradient sensitivity; and a structural term refines the ordering via within-voxel rank, temporal-bin rank, and occupancy sparsity. The three normalized evidence channels are fused into a final VGER score per

What would settle it

Take a new event dataset and backbone, fix all hyperparameters by a blind search that never sees the test set, and run VGER against PCS-xyt with the two-tail protocol; if VGER fails to beat PCS-xyt on both tails in most settings, the reported macro gains are settings-specific rather than a property of the method. A more direct check: replace the center-contraction voxel perturbation in Eq. (10) with an actual deletion of the voxel's events; if the high-tail improvements over PCS-xyt shrink or vanish, the claimed regional evidence is an artifact of the contraction operator rather than of voxel-

Watch

Extended reading notes

Core claim

VGER resolves the granularity dilemma in event attribution by transferring regional voxel evidence to individual events instead of choosing one granularity. It voxelizes the normalized event stream, measures the task-loss change when each voxel's events contract toward the stream center, and broadcasts that loss change to every constituent event. It adds an anisotropic per-event gradient score and a structural ranking term (within-voxel rank, temporal-bin rank, occupancy). Evaluated on three event benchmarks with PointNet, PointNet++, and EventMamba, VGER lowers high-tail and raises low-tail deletion AUCs relative to PCS-xyt in all nine settings.

Load-bearing premise

The ranking's decision-critical localization rests on the premise that moving all events in a voxel toward the stream center while preserving count, polarity, and temporal order produces a loss change that faithfully measures that voxel's true contribution to the prediction; if that perturbational proxy is unfaithful, the voxel channel that the ablation identifies as the main source of critical-event localization would corrupt the ranking.

Editorial extensions

If this is right

  • A single training-free ranking can serve both explanation (which events to inspect or delete to change a prediction) and compression (which events to remove while preserving accuracy).
  • The two-tail deletion protocol gives a unified faithfulness test: high-tail tests prediction-criticality, low-tail tests limited influence, and VGER wins both in all nine reported settings.
  • The anisotropic treatment of temporal displacement relative to spatial displacement ($\rho$) lets the attribution respect the different units of time and space in event streams.
  • The method is backbone-agnostic: it works with global max-pooling (PointNet), hierarchical neighborhoods (PointNet++), and selective state-space models (EventMamba), suggesting event-level attribution is attainable for a broad family of point-based event networks.

Reading between the lines

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

  • The center-contraction perturbation is one specific counterfactual: it moves events toward a robust center rather than deleting them. A natural test is whether an actual removal or outward-shift counterfactual yields the same tail separation; if the gains are specific to the contraction operator, the claimed regional proxy is operator-dependent.
  • The balancing weights are hyperparameters reported per setting. A principled scheme that sets them from model-sensitivity statistics rather than grid search would make the method easier to transfer and reduce the risk that the reported gains are tuning artifacts.
  • The two-tail ranking perspective could extend to insertion curves and to other structured point clouds (e.g., LiDAR), where regional structure also matters; the protocol defines a benchmark that other attribution methods could be measured against.
  • The structural term currently encodes relative sparsity; a semantic grouping (e.g., by object part or motion segment) could replace voxels to capture structure beyond spatio-temporal occupancy.
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 / 4 minor

Summary. The paper proposes VGER, a training-free attribution framework for point-based event cloud networks. VGER combines event-level anisotropic PCS gradient evidence (Eq. 8), task-aware voxel perturbation evidence (Eqs. 9-11), and a structural ranking term (Eqs. 12-17) into a single global event score (Eq. 18). The resulting ordering is evaluated by high-tail and low-tail deletion AUC on three event benchmarks (DVS Gesture, N-MNIST, N-Caltech101) with three backbones (PointNet, PointNet++, EventMamba). The authors report that VGER improves both deletion tails over PCS-xyt in all nine dataset-backbone settings, and an ablation on PointNet++/N-MNIST attributes the high-tail gain primarily to voxel evidence and the low-tail gain to the structural term.

Significance. If the empirical claims hold, VGER would be a valuable contribution to event-based interpretability: it is training-free, addresses the granularity dilemma between point-level and region-level attribution, and introduces a two-tail ranking protocol that jointly tests decision-critical and decision-irrelevant events. The authors also provide an archived evidence trail and reproducible deletion protocol, which is a strength. However, the significance is currently limited by validation concerns: the method has many free hyperparameters, the main text states that setting-specific fusion weights and hyperparameter search results are relegated to an appendix, and the quantitative comparison is made against only one point-level saliency baseline (PCS-xyt) plus Random. Thus the headline claim of consistent improvement across nine settings is not yet firmly established.

major comments (4)
  1. [Experiments / Main Two-Tail Fidelity] The central claim—consistent 9/9 improvement—rests on a ranking score with a dozen free hyperparameters (rho in Eq. 4, alpha and epsilon in Eq. 8, beta_p/beta_v in Eq. 12, w_b/w_c/w_t/w_r/w_d in Eq. 17, lambda_v/lambda_p/lambda_r in Eq. 18, plus grid sizes). The paragraph after Table 2 states that the appendix contains 'setting-specific fusion weights used in part of the experiments' and 'auxiliary results from hyperparameter search.' If these weights were selected on the same test sets used to compute Table 2, the reported consistency could arise from per-setting overfitting rather than from a genuinely better ranking. The manuscript must (a) describe the validation protocol used to select weights, (b) report the actual weights for all nine settings in the main text, and (c) provide a sensitivity analysis or show that fixed hyperparameters across settings preserve the qualitative result
  2. [Table 2 and Ablation Study] No error bars, confidence intervals, or multiple model seeds are reported. Each AUC in Table 2 is a single number from a frozen model; for deep networks, training stochasticity can materially change attribution rankings and deletion curves, especially on small test sets. The ablation is performed only on one development subset (PointNet++/N-MNIST), so the claim that voxel evidence is the main source of high-tail localization and the structural term improves the low tail is supported by only one setting. Please report results over at least three model seeds per setting, or otherwise justify that single-model results are stable.
  3. [Experiments / Baselines] The abstract and introduction claim improvement 'over point-level saliency baselines' (plural), but the quantitative comparison in Tables 1-2 includes only PCS-xyt and Random. Basic Voxel is mentioned in a separate PointNet study, not in the main 3x3 matrix. To support the plural claim and to show VGER is not merely better than one baseline, the authors should evaluate at least one or two additional point-level attribution methods (e.g., gradient saliency, Integrated Gradients, or a PCS variant) under the same two-tail protocol, or revise the wording to 'over PCS-xyt and Random.'
  4. [Eq. (10) and Eq. (11)] The voxel evidence in Eq. (10) is computed by moving all events in a voxel toward the stream center, not by deleting them, whereas the evaluation protocol deletes events. The assumption that this local-perturbation proxy faithfully identifies decision-critical events is asserted but not directly validated. The ablation shows a positive correlation on one development setting, but this does not establish that the proxy is reliable across architectures, datasets, and deletion ratios. Please add a concrete validation: e.g., compare voxel evidence obtained by center-shifting against voxel evidence obtained by actual deletion on a development subset, or report the effect of varying the shift magnitude.
minor comments (4)
  1. [Eq. (8)] The statement that Eq. (8) 'reduces to the original isotropic PCS formulation when rho=1' needs a precise reference to Zheng et al. 2019 and an explicit derivation; the role of alpha and epsilon should be stated.
  2. [Eq. (10)] The perturbation eE(j) is not fully defined: how far are events moved, in what direction, and how are boundary/tie cases handled? Please specify the construction formally.
  3. [Problem Formulation] The deletion operator D is mentioned but not defined precisely; clarify whether deletion is performed on the raw event set or on the backbone input representation, and how polarity and timestamps are handled by each backbone.
  4. [Figure 3] The legend lists five curves; please make the line styles/colors distinguishable in grayscale print and provide a short description of the low-tail/high-tail construction in the caption.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: VGER's ranking is not equivalent to its evaluation by construction; the only concern is undisclosed hyperparameter selection, which is a validation risk, not a circular derivation.

full rationale

VGER's final score (Eq. 18) is a weighted combination of three independent signals: event-level anisotropic PCS gradients (Eq. 8), voxel-level task loss change from a move-to-center perturbation (Eq. 10), and a structural rank term (Eq. 17). The evaluation (Eq. 2 and the two-tail deletion protocol) removes top/bottom events and measures accuracy, which is a different operation from the move-to-center perturbation used in attribution; the score is not defined in terms of deletion AUC, so the reported high/low tail separation is not forced by definition. The voxel evidence is broadcast to events (Eq. 11) but this is a resolution transfer, not a circular definition. PCS-xyt is a component of VGER, but the method adds task-aware voxel evidence and structural ranking, and the gains over PCS-xyt are empirical rather than algebraic. The self-citations (EventMamba backbone, related point-based event network papers) are not load-bearing for the derivation: EventMamba is used as an external benchmark, and no uniqueness theorem or prior-work premise is invoked to rule out alternatives. The only flagged concern is the appendix note about 'setting-specific fusion weights' and 'hyperparameter search results' with no explicit validation protocol in the supplied text; if those weights were tuned on the same test AUCs, the consistency claim would be a model-selection artifact. However, the text does not establish that, per-setting hyperparameters are normal practice, and this is a correctness/transparency risk rather than a circularity. The derivation chain from Eqs. (3)-(19) is self-contained.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

VGER introduces no new physical entity, but it relies on many hand-set fusion weights and on a custom voxel perturbation proxy. The number of free parameters is large relative to the evidence provided, and the appendix with their values and selection procedure is not included.

free parameters (7)
  • rho (anisotropic temporal weight) = unknown
    Controls relative weighting of temporal vs spatial displacement in Eqs. (4)-(8); no value reported in main text.
  • alpha (radial scaling exponent) = unknown
    Controls radial scaling in the anisotropic PCS score, Eq. (8); needed to compute the gradient evidence.
  • beta_p and beta_v (base fusion weights) = unknown
    Weights in Eq. (12) that combine normalized gradient and voxel evidence into the base score.
  • w_b, w_c, w_t, w_r, w_d (structural weights) = unknown
    Weights in Eq. (17) for base score, voxel rank, temporal rank, sparsity, and occupancy; jointly define the structural term.
  • lambda_v, lambda_p, lambda_r (final fusion weights) = unknown
    Weights in Eq. (18) that fuse voxel, gradient, and structural evidence into the final VGER score.
  • G_s and G_t (voxel grid sizes) = unknown
    Number of spatial and temporal voxels used to compute regional perturbation evidence; not reported in the main text.
  • epsilon (stability constant) = unknown
    Small constant in Eq. (8) to avoid division by zero; minor but unspecified.
assumptions (5)
  • domain assumption Gradients of the frozen network's task loss with respect to normalized event coordinates are meaningful and computable for PointNet, PointNet++, and EventMamba.
    The gradient evidence in Eq. (8) requires differentiable access through the backbone and task loss; not all event models are differentiable in this way.
  • ad hoc to paper Moving all events in a voxel toward the stream center, while preserving count, polarity, and temporal order, is a faithful local perturbation whose loss change measures the region's contribution.
    This is the core regional proxy in Eq. (10). If this perturbation does not approximate removal or corruption of the region, the voxel evidence is not valid.
  • ad hoc to paper Broadcasting voxel loss change to each constituent event (Eq. 11) preserves event-level evidence and is a sound transfer operation.
    This transfer is the paper's key bridge between regional and event granularity, but no proof or empirical validation of its soundness is given.
  • domain assumption Fixed-cardinality deletion curves are a valid ground-truth measure of attribution faithfulness.
    The paper cites deletion/insertion literature and acknowledges caveats via Hooker et al. and Tomsett et al., yet uses deletion AUC as the primary fidelity criterion.
  • ad hoc to paper Stream-wise score normalization does not distort cross-event comparisons.
    Norm is applied independently to each evidence stream in Eqs. (12) and (18); the effect of this choice on the final ranking is not analyzed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VGER: Voxel-Guided Global Event Ranking for Event Cloud Attribution." pith.science (2026). https://pith.science/paper/P6ECPX6Z

@misc{pith2026260801470,
  author       = {Pith},
  title        = {Pith review of: VGER: Voxel-Guided Global Event Ranking for Event Cloud Attribution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P6ECPX6Z}},
  note         = {Machine review of arXiv:2608.01470}
}
read the original abstract

Event cameras produce sparse and asynchronous event streams that provide rich spatio-temporal information for efficient perception. Recent advances in event-based models have demonstrated strong performance by directly modeling asynchronous events without dense frame reconstruction. However, identifying the event-level evidence behind their predictions is crucial for improving model transparency and reliability. Directly adapting point-level saliency methods from point clouds provides fine-grained attribution but overlooks event-specific spatio-temporal structures. To address this limitation, we propose Voxel-Guided Global Event Ranking (VGER), a training-free attribution framework for point-based event cloud networks. VGER combines event-level gradient evidence with task-aware voxel perturbation evidence, transferring regional contribution into event-level attribution scores while preserving fine-grained resolution. Furthermore, VGER introduces a unified event ranking strategy, where high-ranked events are expected to be prediction-critical and low-ranked events are expected to have limited influence on predictions. We evaluate VGER on three event-based benchmarks with PointNet, PointNet++, and EventMamba. Across nine dataset-backbone settings, VGER consistently improves both high-tail and low-tail deletion performance over point-level saliency baselines.

Figures

Figures reproduced from arXiv: 2608.01470 by the authors.

Figure 1
Figure 1. Event-level saliency maps produced by [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of VGER. Event-level gradient evidence and task-aware voxel perturbation evidence are transferred, normalized, and fused into one global event ordering. The same ordering is evaluated at the high and low tails to distinguish prediction-critical events from events with limited influence. Problem Formulation We consider the attribution problem for point-based event cloud networks. An event stream is represent… view at source ↗
Figure 3
Figure 3. Two-tail deletion curves for the 3×3 matrix of backbones and datasets. Each panel deletes 5%–90% of events in score order and reports accuracy. VGER’s high tail (VGER-H) drops below PCS-xyt-H and Random in most settings and ratios, and its low tail (VGER-L) stays above PCS-xyt-L in most settings and ratios, so the two VGER tails separate decision-critical from decision-irrelevant events across the whole protocol. Re… view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Representative prediction changes after delet [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 25 canonical work pages

  1. [1]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    Event-based vision: A survey , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2022 , publisher=

    work page 2022

  2. [2]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    End-to-end learning of representations for asynchronous event-based data , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

    work page

  3. [3]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    HATS: Histograms of averaged time surfaces for robust event-based object classification , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    work page

  4. [4]

    a sparse-aware solution for efficient event data processing , author=

    Event transformer. a sparse-aware solution for efficient event data processing , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  5. [5]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Get: Group event transformer for event-based vision , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    work page

  6. [6]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Pointnet: Deep learning on point sets for 3d classification and segmentation , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  7. [7]

    Advances in neural information processing systems , volume=

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space , author=. Advances in neural information processing systems , volume=

  8. [8]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    Rethinking efficient and effective point-based networks for event camera classification and regression , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    work page

Show all 50 references
  1. [9]

    Advances in Neural Information Processing Systems , volume=

    E2pnet: event to point cloud registration with spatio-temporal representation learning , author=. Advances in Neural Information Processing Systems , volume=

  2. [10]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    A simple and effective point-based network for event camera 6-dofs pose relocalization , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  3. [11]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Pointcloud saliency maps , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  4. [12]

    arXiv preprint arXiv:1312.6034 , year=

    Deep inside convolutional networks: Visualising image classification models and saliency maps , author=. arXiv preprint arXiv:1312.6034 , year=

    arXiv

  5. [13]

    International conference on machine learning , pages=

    Axiomatic attribution for deep networks , author=. International conference on machine learning , pages=. 2017 , organization=

    2017

  6. [14]

    Proceedings of the IEEE international conference on computer vision , pages=

    Grad-cam: Visual explanations from deep networks via gradient-based localization , author=. Proceedings of the IEEE international conference on computer vision , pages=

  7. [15]

    arXiv preprint arXiv:1806.07421 , year=

    Rise: Randomized input sampling for explanation of black-box models , author=. arXiv preprint arXiv:1806.07421 , year=

    arXiv

  8. [16]

    Proceedings of the IEEE international conference on computer vision , pages=

    Interpretable explanations of black boxes by meaningful perturbation , author=. Proceedings of the IEEE international conference on computer vision , pages=

  9. [17]

    Advances in neural information processing systems , volume=

    Sanity checks for saliency maps , author=. Advances in neural information processing systems , volume=

  10. [18]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Sanity checks for saliency metrics , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  11. [19]

    Advances in neural information processing systems , volume=

    On the (in) fidelity and sensitivity of explanations , author=. Advances in neural information processing systems , volume=

  12. [20]

    Advances in neural information processing systems , volume=

    A benchmark for interpretability methods in deep neural networks , author=. Advances in neural information processing systems , volume=

  13. [21]

    International Conference on Pattern Recognition and Artificial Intelligence , pages=

    Metrics for saliency map evaluation of deep learning explanation methods , author=. International Conference on Pattern Recognition and Artificial Intelligence , pages=. 2022 , organization=

    2022

  14. [22]

    arXiv preprint arXiv:2403.07706 , year=

    Fast and simple explainability for point cloud networks , author=. arXiv preprint arXiv:2403.07706 , year=

    arXiv

  15. [23]

    arXiv preprint arXiv:2410.15374 , year=

    Explainability of point cloud neural networks using smile: Statistical model-agnostic interpretability with local explanations , author=. arXiv preprint arXiv:2410.15374 , year=

    arXiv

  16. [24]

    IEEE Access , year=

    XAI for Point Cloud Data Using Perturbations Based on Meaningful Segmentation , author=. IEEE Access , year=

  17. [25]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

    Eventpoint: Self-supervised interest point detection and description for event-based camera , author=. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

  18. [26]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Interpretable3d: An ad-hoc interpretable classifier for 3d point clouds , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  19. [27]

    Attention is not explanation , author=. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) , pages=

    2019

  20. [28]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    A low power, fully event-based gesture recognition system , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  21. [29]

    Frontiers in neuroscience , volume=

    Converting static image datasets to spiking neuromorphic datasets using saccades , author=. Frontiers in neuroscience , volume=. 2015 , publisher=

    2015

  22. [30]

    Frontiers in Neuroscience , volume=

    Gradient-based feature-attribution explainability methods for spiking neural networks , author=. Frontiers in Neuroscience , volume=. 2023 , publisher=

    2023

  23. [31]

    Advances in Neural Information Processing Systems , volume=

    FFAM: feature factorization activation map for explanation of 3d detectors , author=. Advances in Neural Information Processing Systems , volume=

  24. [32]

    arXiv preprint arXiv:2606.23078 , year=

    A Systematic Survey on Event Camera Representation Learning , author=. arXiv preprint arXiv:2606.23078 , year=

    arXiv

  25. [33]

    European Conference on Computer Vision , pages=

    A differentiable recurrent surface for asynchronous event-based data , author=. European Conference on Computer Vision , pages=. 2020 , organization=

    2020

  26. [34]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Eventnet: Asynchronous recursive event processing , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  27. [35]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Aegnn: Asynchronous event-based graph neural networks , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  28. [36]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    A voxel graph cnn for object classification with event cameras , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  29. [37]

    Forty-third International Conference on Machine Learning , year=

    Scalable Event Cloud Network for Event-based Classification , author=. Forty-third International Conference on Machine Learning , year=

  30. [38]

    arXiv preprint arXiv:2602.08230 , year=

    Generating Adversarial Events: A Motion-Aware Point Cloud Framework , author=. arXiv preprint arXiv:2602.08230 , year=

  31. [39]

    Why should i trust you?

    " Why should i trust you?" Explaining the predictions of any classifier , author=. Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining , pages=

  32. [40]

    Advances in neural information processing systems , volume=

    A unified approach to interpreting model predictions , author=. Advances in neural information processing systems , volume=

  33. [41]

    PloS one , volume=

    On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation , author=. PloS one , volume=. 2015 , publisher=

    2015

  34. [42]

    arXiv preprint arXiv:1706.03825 , year=

    Smoothgrad: removing noise by adding noise , author=. arXiv preprint arXiv:1706.03825 , year=

    arXiv

  35. [43]

    2018 IEEE winter conference on applications of computer vision (WACV) , pages=

    Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks , author=. 2018 IEEE winter conference on applications of computer vision (WACV) , pages=. 2018 , organization=

    2018

  36. [44]

    European conference on computer vision , pages=

    Visualizing and understanding convolutional networks , author=. European conference on computer vision , pages=. 2014 , organization=

    2014

  37. [45]

    Explainable AI: Interpreting, explaining and visualizing deep learning , pages=

    The (un) reliability of saliency methods , author=. Explainable AI: Interpreting, explaining and visualizing deep learning , pages=. 2019 , publisher=

    2019

  38. [46]

    arXiv preprint arXiv:1806.08049 , year=

    On the robustness of interpretability methods , author=. arXiv preprint arXiv:1806.08049 , year=

    arXiv

  39. [47]

    ACM Transactions on Graphics (tog) , volume=

    Dynamic graph cnn for learning on point clouds , author=. ACM Transactions on Graphics (tog) , volume=. 2019 , publisher=

    2019

  40. [48]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Point transformer , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  41. [49]

    Frontiers in neuroscience , volume=

    Cifar10-dvs: an event-stream dataset for object classification , author=. Frontiers in neuroscience , volume=. 2017 , publisher=

    2017

  42. [50]

    International Conference on Learning Representations , volume=

    Spikepoint: An efficient point-based spiking neural network for event cameras action recognition , author=. International Conference on Learning Representations , volume=

Pith tools

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