acceptodds
Withdrawn from ICLR 2027

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

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Event cameras produce sparse, asynchronous streams in which spatial coordinates and timestamps play fundamentally different physical and semantic roles. Consequently, saliency methods developed for static 3D point clouds do not transfer naturally to event point clouds. We introduce VGER, a post-hoc event-level attribution method for point-based event recognition models. VGER combines anisotropic gradient evidence, a task-aware voxel prior, and local rank calibration. Its anisotropic formulation models spatial, temporal, and joint spatiotemporal sensitivity separately instead of treating normalized spatial and temporal coordinates as geometrically interchangeable. Changes in task loss under voxel perturbations identify relevant spatiotemporal regions, while local spatial and temporal ranks distribute this regional evidence among individual events. The resulting global score requires no model retraining and supports a unified two-tail evaluation: removing highly ranked events should rapidly degrade recognition performance, whereas removing low-ranked events should preserve it. We evaluate VGER using a proportional-deletion protocol on DVS Gesture, N-MNIST, and N-Caltech101 with PointNet, PointNet++, and EventMamba backbones. The evaluation examines attribution fidelity, stability, robustness, cross-backbone consistency, and computational efficiency.

Primary area:
interpretability and explainable AI
Keywords:
event-based vision, event cameras, event cloud networks, explainable AI

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Pn28OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the attribution task for point‑based event‑camera networks. Existing methods mainly fall into two categories: point‑level‑gradient‑based methods and voxel‑perturbation‑based methods. Both paradigms suffer from certain limitations. Specifically, the former lacks task‑driven perturbation evidence for local spatio‑temporal regions, while the latter fails to distinguish individual events within each voxel. This work proposes to fuse these two paradigms and introduces a two‑tail deletion evaluation protocol. Experiments are conducted on three event‑classification datasets with three backbone networks. However, it does not perform comparisons against existing state‑of‑the‑art methods.

Critical strengths

The authors gain insight into the strengths and weaknesses of existing paradigms as well as their complementarity, and accordingly propose combining the merits of the two paradigms to mitigate their respective limitations.

Critical weaknesses
  1. This paper lacks comparisons against state‑of‑the‑art methods.
  2. The attribution is conditioned on ground‑truth labels rather than the model’s predicted outputs.
  3. When computing voxel contributions, events are shifted toward the stream center for replacement instead of being truly deleted. By contrast, direct event deletion is adopted in the evaluation phase. These two intervention operations are not equivalent.
Questions for authors
  1. The structural ranking term re‑uses the base score that already fuses gradient and voxel signals, feeding the evidence repeatedly into the fusion pipeline. What is the motivation behind such a design?
  2. Voxel‑level perturbation requires additional model forward passes for every occupied spatio‑temporal voxel. Compared with gradient‑only attribution methods, it brings a substantial increase in computational cost. How can this limitation be mitigated when fusing the two paradigms?
Reviewer_3N2zOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose VGER, a training-free attribution method for point-based event cloud networks, with a fused score that combines PCS-style event gradients, voxel perturbation loss changes, and a hand-designed structural ranking term. Experiments on DVS Gesture, N-MNIST, and N-Caltech101 with PointNet, PointNet++, and EventMamba demonstrate lower high-tail and higher low-tail deletion AUC than PCS-XYT and Random.

Critical strengths
  1. The central idea is intuitive and well motivated. Transferring voxel-level evidence back to individual events is a reasonable way to bridge these two granularities, and the formulation is clear enough to reproduce.
  2. The method is training-free and largely model-agnostic. It only needs input gradients and forward passes through a frozen network, so it can be applied to different point-based backbones without architectural changes or extra training.
Critical weaknesses
  1. The results are internally inconsistent. The main text says VGER improves both tails in all nine settings, but the appendix contradicts this. For PointNet++/DVS Gesture, Table 2 reports a high-tail AUC of 0.4665 vs 0.5311 for PCS-XYT and a low-tail AUC of 0.6035 vs 0.5315. Table 6 gives 0.5492 vs 0.5303 and 0.5107 vs 0.5313 for the same comparisons. By Table 6, VGER loses both tails in this setting and also falls behind Random (0.5244). For EventMamba/N-Caltech101, all four entries differ: VGER-H is 0.2180 in Table 2 but 0.1323 in Table 11, and VGER-L is 0.3381 vs 0.2399. The authors acknowledge these discrepancies but leave them unresolved, which undermines confidence in the reported gains.
  2. The evaluation is confounded by label information. Both the attribution scores and the evaluation metric depend on the ground-truth label. As a result, low-tail deletion can remove counter-evidence and raise accuracy above the unperturbed level. The low-tail metric therefore does not measure faithfulness to the model's own decision.
  3. It is unclear where the improvement comes from. The structural rank term may help the low tail simply by spreading deletions evenly across voxels, which acts like stratified subsampling. The ablation does not isolate this effect, and a stratified random baseline is missing.
Questions for authors

Please refer to Weakness part.

Reviewer_sLYfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents VGER, which combines event-level gradient attribution, voxel-level perturbation, and some structural ranking rules to explain point-based event-camera classifiers. The idea of fusing regional and event-level info is intuitive, and the authors test it on three datasets and three architectures. However, there are contradictory numbers in the experiments, and a mismatch between what they actually measure and how they interpret it, which makes it hard to really trust how faithful the attribution is.

Critical strengths
  1. The explain for sparse spatiotemporal event classifiers is a relevant problem.
  2. Mixing regional perturbations with single-event sensitivity makes sense, at the same time, it can keep the output at the event level.
Critical weaknesses
  1. Table 2 and Tables 6/11 have large inconsistency. Table 6 goes against the claim that the method improves on both tails in every setting.
  2. The evaluation setup and how events are modified are not consistent. The main text says they do full-range deletion, but the tables show normalized 5%–90% AUC. Also, the voxel scoring replaces events with the center, while the evaluation just deletes them.
  3. Low-ranked events can be evidence pushing the model toward a wrong answer. Removing them can improve accuracy precisely because they were influential, not irrelevant. Accuracy after removal therefore does not justify the claims of decision irrelevance or prediction-preserving compression.
Questions for authors

See above.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.