acceptodds
Withdrawn from ICLR 2027

Exchange, Not Gating: Event-Guided Deblurring and Deraining

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Event cameras record brightness changes asynchronously at high temporal resolution, making them a natural guide for RGB restoration tasks such as motion deblurring and image deraining. Existing methods, however, still couple the two modalities through three heuristic choices: they process event voxels with blocks designed for images, control event injection with hand-designed activity masks and unconstrained gates, and decode fused features without directional guidance. We replace these choices with three complementary mechanisms. VIBE provides event-native processing by separating polarities and conditioning the event stream on the mass, signed mass, temporal centroid and temporal spread of each pixel's event histogram, thereby removing the baseline's hand-designed blur-and-mask pipeline. COEX replaces independent gates with a per-channel mixer projected toward the doubly stochastic manifold, encouraging convex cross-modal mixing while approximately preserving the sum of the two inputs. COMPASS predicts a per-pixel distribution over five fixed shifts from event features and their moment condition, realigning fused features without regressing offsets. The resulting single-frame model achieves the best event-guided results on GoPro, HighREV and NTURain: on GoPro, it reaches dB compared with dB from the official TRM-UNet weights, while a larger configuration reaches dB.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
event camera, image restoration, motion deblurring, image deraining

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_svriOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes three modules, VIBE, COEX, and COMPASS, for event-guided image deblurring and deraining. VIBE augments event features with several hand-crafted temporal statistics, COEX replaces conventional fusion gates with a projected 2 x 2 mixing matrix, and COMPASS uses event-conditioned weights over five fixed spatial shifts. The resulting model reports strong numbers on GoPro, HighREV, and NTURain.

Critical strengths
  • The paper does evaluate on both synthetic and real-event settings, and the reported numerical performance is strong if the comparisons are correct. In particular, the method reports improvements on GoPro, HighREV, and NTURain rather than relying on a single benchmark.
  • The authors at least attempt controlled component studies rather than reporting only a full-model result. The retrained GoPro baseline and several replacement experiments are useful in principle, although I do not think the current ablations are sufficient to support the stronger mechanism-level claims.
Critical weaknesses

Before the technical issues, I have concerns about the overall level of care. Figure 1 does not clearly communicate where VIBE, COEX, and COMPASS operate or what their defining computations are, and the later figures are similarly cramped and difficult to inspect. The Related Work section also mixes equations and dataset statistics into the literature discussion, while notation such as G_\sigma remains undefined even when reused in Sec. 3.1. Given the authors’ disclosed use of generative AI for manuscript drafting, analysis and plotting scripts, literature surveying, and discussion of the ablation matrix, I would have expected substantially more careful human verification and editing.

  1. The motivation appears somewhat reverse-engineered around the three proposed modules. The paper summarizes existing event-guided restoration through three “heuristic design choices”: image-like event processing, gating-based fusion, and direction-agnostic decoding. This characterization overlooks recent methods that already use motion-aware event representations, asymmetric cross-modal interaction, and explicit edge or motion modeling [a–d]. The proposed modules themselves also feel rather ad hoc. VIBE uses four hand-selected statistics, COEX imposes an insufficiently justified conservation constraint on learned latent features, and COMPASS is essentially an event-conditioned five-tap dynamic filter. The additional complexity of COEX is particularly unclear given that a 2x2 doubly stochastic mixer can be parameterized much more simply.
  2. I have more serious concerns about the experimental comparison. Table 1 reports MAENet at 34.50 dB / 0.9534 on GoPro, whereas [a] reports 36.07 dB. AHDINet is similarly reported as 36.98 / 0.9642, compared with 37.09 / 0.9820 in [b]. These discrepancies are too large to dismiss as rounding, and the provenance of the baseline numbers is not sufficiently clear. The comparison also omits highly relevant recent methods [c,d]. In addition, the ablations are conducted only on synthetic-event GoPro and are mostly cumulative, so they do not establish that VIBE explains the real-event improvement on HighREV or that COMPASS provides the claimed directional benefit on NTURain. The modified EGVD evaluation is also difficult to interpret because its recurrent state is reset for every input after the standard setting reportedly “collapses.”
  • References
    • [a] Motion Aware Event Representation-driven Image Deblurring, ECCV 2024.
    • [b] Asymmetric Hierarchical Difference-aware Interaction Network for Event-guided Motion Deblurring, AAAI 2025.
    • [c] Separation for Better Integration: Disentangling Edge and Motion in Event-based Deblurring, ICCV 2025.
    • [d] Event-Based Motion Deblurring Using Task-Oriented 3D Gaussian Event Representations, CVPR 2026.
Questions for authors
  • Please clarify the provenance of the baseline results in Tables 1–3. Which numbers are taken from the original papers, which are obtained by reevaluating official checkpoints, and which are produced by the authors’ own retraining? In particular, the reported MAENet and AHDINet results on GoPro appear inconsistent with their published numbers. These discrepancies need to be clarified before the claimed SOTA improvements can be properly assessed.
  • The paper organizes the motivation around three supposedly common limitations of existing event-guided restoration methods. How broadly do these three limitations actually apply to recent methods? In particular, how are the proposed ideas positioned against “Motion Aware Event Representation-driven Image Deblurring” (ECCV 2024), “Separation for Better Integration: Disentangling Edge and Motion in Event-based Deblurring” (ICCV 2025), and “Event-Based Motion Deblurring Using Task-Oriented 3D Gaussian Event Representations” (CVPR 2026)? The latter recent works also seem highly relevant to the claimed event-specific representation and motion-aware modeling.
  • The component analysis is conducted only on GoPro. Can the authors provide at least a basic ablation on HighREV and NTURain? Since the paper motivates VIBE partly through event reliability and COMPASS through directional degradation, it would be important to verify that these components remain useful on real events and on deraining rather than only on synthetic-event GoPro.
  • Why is EGVD evaluated with its recurrent state reset for every input? Since temporal recurrence is part of the original method, this modification appears to change its inference behavior. What happens under the original sequential inference protocol, and why does that setting “collapse” in the authors’ reproduction?
  • The real-event deblurring evaluation is currently limited to HighREV. Why are other commonly used real-event benchmarks, such as REBlur, not included? Given the paper’s relatively broad claims about event-guided deblurring, additional real-event evaluation would make the empirical conclusions substantially more convincing.
Reviewer_tvaSOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

To better utilize event information, this paper proposes three complementary mechanisms—VIBE, COEX, and COMPASS—for event-guided image restoration. The authors conduct extensive experiments on the GoPro, HighREV, and NTURain datasets to demonstrate the effectiveness of the proposed method.

Critical strengths

Proposed method achieve SOTA performance across three different tasks on both synthetic and real datasets. The proposed method also have relative high efficiency.

Critical weaknesses
  1. There is a lack of visualziation comparsion for the ablation studies. It is hard to support the effectiveness of proposed modules with only numerical results.
  2. Although HighREV dataset contains real events, most of datasets are systhntic, escpailly for GoPro with both sysntethic degradation and events. Authors should consider adding more experiments in real-world datasets to better demonstrates the robustness of the proposed method.
  3. It is hard to identify the motivation or evidence for the designed modules. Polarity-separated event representations have been used before, while COEX and COMPASS lack convincing event-specific giagonistics.
  4. It is hard to identify VIBE, COEX, or COMPASS, or see how the moment condition, mixing matrix, and directional weights are generated in Fig.1.
Questions for authors

Please address the concerns in Critical Weaknesses.

Reviewer_Uv2COpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies event-assisted image deblurring and deraining, using bidirectional interaction rather than gating to make fuller use of event information. VIBE uses event polarity and temporal statistics to construct event representations; COEX performs feature interaction through an approximately doubly stochastic matrix; COMPASS predicts local weights from event features to guide reconstruction during decoding. The authors validate the effectiveness of the method on benchmark datasets.

Critical strengths

The paper achieves improvements on both deblurring and deraining tasks, with evaluations covering synthetic and real event data. In addition, the authors progressively add components to demonstrate their effectiveness.

Critical weaknesses

1.The paper centers on moving from gating to interaction, but the experiments do not demonstrate failure cases of other baseline designs. The current experiments only verify quantitative improvements after adding the proposed components, without explaining the limitations of the original ways of using event information or why interaction can address them.

2.COEX is essentially bidirectional soft gating with complementary weights, and it is not fundamentally distinct from gating. Table 4 shows that removing the projection reduces performance, indicating that the projection is helpful. However, does the benefit come from weight normalization and mixing, or from the constraints on row and column sums?

3.The current COMPASS experiments only demonstrate that spatial adaptation outperforms global processing. However, the authors need to explain and demonstrate why the weights should be predicted from the event side rather than from RGB or fused features during decoding. How can spatially varying weights be shown to correspond to motion directions?

Questions for authors

1.How do existing gating methods process and use event information? Comparisons are needed to show that VIBE provides improvements.

2.Additional evidence is needed to distinguish the effects of ordinary normalization from those of the row- and column-sum constraints in COEX.

3.It is necessary to verify that predicting weights from event-side features in COMPASS offers advantages over predicting them from RGB or fused features. Furthermore, how can the correspondence between its spatial weights and local motion or rain-streak directions be demonstrated?

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

Related papers

Discussion (0)

Sign in to comment.