EvFocus: Learning to Reconstruct Sharp Images from Out-of-Focus Event Streams

We sincerely thank the reviewer for the detailed and thoughtful feedback, and we carefully address the raised concerns below.

---

Realistic 3D scenes and evaluation

Thank you for pointing this out. We have conducted additional experiments on realistic 3D scenes with varying depth levels. The qualitative results are shown in the updated figure (https://anonymous.4open.science/api/repo/Anonymous-6BB9/file/fig1.pdf?v=9bbf2ec6).

The results demonstrate that our model effectively learns the relationship between event streams at different defocus levels and corresponding illumination intensities. It is capable of handling varying degrees of defocus blur across different depth planes in the scene.

In addition, we conducted quantitative evaluations on these 3D scenes using standard no-reference image quality metrics:

These results indicate that our model consistently achieves better sharpness and contrast while reducing noise and preserving perceptual quality. We will incorporate these new results, the introduction of metrics, and the analysis in the revised version.

---

​Depth-aware feature maps & metrics

We appreciate the reviewer’s suggestions. We agree that PSF-aware validation could provide deeper insight into the model's physical consistency. However, similar to most image-based deblurring networks, our model does not include an explicit PSF estimation module, which makes such direct evaluation difficult.

Thus, to evaluate the model’s performance on multi-depth defocus removal, we performed feature visualizations on 3D scenes. In the updated figure (https://anonymous.4open.science/api/repo/Anonymous-6BB9/file/fig2.pdf?v=72d4d619), the result demonstrates that our model is able to distinguish between objects near and far from the focal plane, indicating that it learns depth-aware defocus representations implicitly through training. In the revised version, we will also report video-level metrics such as frame-wise stability and temporal flicker scores to better reflect temporal behavior.

---

Visual quality and artifacts

Thank you for pointing this out. We acknowledge that some visual artifacts remain, especially in challenging areas like edges or low-texture regions. However, this task is inherently difficult; the defocused event streams are sparse, noisy, and lack full spatial information, making accurate reconstruction very challenging. Despite this, our method achieves significantly better perceptual quality than existing baselines, as shown by multiple no-reference metrics in the tables. We will add more 3D scene results in the revised version.

---

Model and Dataset Design

Thank you for the comments.
Our synthetic dataset is built on a physically-inspired defocus model that simulates event generation under varying PSF sizes and lighting conditions. By changing blur levels, motion speed, and intensity across sequences, we mimic different scene depths and real-world defocus effects. This design allows the model to learn both temporal accumulation and spatial defocus patterns, and to adapt to the link between blur size and event sparsity, making it effective for depth-varying defocus.

We conducted experiments on multi-depth scenes and observed that E2VID pretrained on sharp data can handle slight defocus, suggesting that events exhibit robustness to slight defocus blur. Combined with our specially designed dataset, our model further learns depth-dependent defocus patterns and can distinguish and reconstruct objects at different depths, showing strong generalization to 3D scenarios.

---

Related work

Thank you for this suggestion. We will discuss relevant references (e.g., Lou et al., 2023; Lin et al., 2022; Teng et al., 2024) on event-driven autofocus and focal stack fusion. Our method differs in that it does not require active lens control or RGB images, and is designed for passive, single-shot defocused event streams, which we believe complements autofocus-based methods.

---

Thank you again for your valuable feedback. If there are any additional questions, we would be happy to address them.

We sincerely thank the reviewer for the encouraging feedback and constructive suggestions. We address the main points below.

---

Additional baseline: event-to-video + image/video deblurring

We appreciate this valuable suggestion. In the supplementary material (Fig. 11–12), we included an experiment that follows a similar two-stage baseline: we first reconstruct intensity images from defocused events, followed by applying NRKNet [CVPR'23] as the SIDD network for defocus deblurring.

However, the results were suboptimal, mainly due to the low quality of reconstructed images from sparse defocused events, which contain significant artifacts and structural degradation.

These observations support our decision to address defocus directly in the event domain, which our method handles more effectively in terms of both reconstruction fidelity and robustness. We will add more discussion in the revised version.

---

Clarification on Brisque and NIQE metrics

Thank you for pointing this out. We will add the following clarifications in the final version:

• BRISQUE (Blind/Referenceless Image Spatial Quality Evaluator) evaluates natural scene statistics in the spatial domain without requiring a reference image.
• NIQE (Natural Image Quality Evaluator) computes quality based on deviations from learned statistical regularities of natural images.

These no-reference metrics are particularly useful for assessing real-world reconstruction results where ground truth is unavailable.

---

Discussion of related work: Lou et al. (CVPR 2023)

Thank you for highlighting this relevant work. Lou et al. [CVPR'23] propose an all-in-focus imaging method based on both RGB images and event focal stacks, which relies on capturing event streams under different focal settings and performing temporal fusion.

In contrast, our work focuses on single-shot defocused event streams, which:

• Require no active focus control or RGB supervision;
• Address the case where defocus has already occurred, aiming to restore sharp content purely from event data.

We will incorporate a discussion of this difference and clarify the complementary nature of the two approaches in the final revision.

---

Runtime comparison

Thanks for your suggestion. We evaluated inference speed and model size under identical settings (input: [1, 5, 264, 352], GPU: RTX 3090). The results are as follows:

Our method offers a balance between performance and complexity. While not the fastest, EvFocus yields consistently higher-quality reconstructions.

---

Extension to changing focal planes

Thanks for your advice. This is an excellent and challenging direction. Our current method assumes static defocus, but we believe it can be extended to handle dynamic focal planes by:

• Incorporating temporal modulation or adaptive kernel modeling;
• Simulating changing focal depth in training data;
• Leveraging the temporal encoder to track evolving blur distributions.

We consider this a promising line of future work and greatly appreciate the suggestion.

---

Thank you again for your thoughtful review and support. We are happy to clarify further if any questions remain.

We thank the reviewer for their thoughtful comments and constructive suggestions. We address each of the raised points below.

---

Computational Efficiency

We agree that runtime performance is important for real-world deployment. In response, we have conducted a comprehensive runtime and parameter comparison (input size: [1, 5, 264, 352] on RTX 3090 GPU):

Our model achieves a good trade-off between accuracy and complexity. The inference time is similar to ET-NET; our method offers significantly improved image quality.

---

Why isn't the real defocus ground truth (GT) used during refocusing?

Thank you for this insightful question. Our model architecture includes a blur-aware dual-branch decoder, where the blur decoder passes through three CF modules and fuses with features from the align decoder. We employ a deblurring reconstruction module on the align path to enhance its output. To further leverage the design, we introduce a loss between the blur decoder output and the re-defocus module to encourage CF modules to extract meaningful blur representations.

One of our intentions is to encourage effective feature interaction between the two branches. Specifically, the blur decoder extracts blur-aware event representations, which are used to guide and support the learning process of the align decoder. This interaction allows the align decoder to better capture defocus-related features from the event domain. In fact, there is a modality gap between events and RGB images: some low-texture regions or subtle motion details present in the GT image may not be well represented in event streams. In this context, directly supervising with defocus GT could potentially lead the optimization to overly favor one branch (e.g., the blur decoder), focusing more on image-domain appearance rather than learning the common defocus patterns reflected in the event stream. This may weaken the collaborative learning dynamics between the two branches.

Our current design aims to balance the contributions of both branches and promote cooperative learning, and has proven effective in practice (validated through ablation studies in the paper). We appreciate the reviewer’s question, and we will include a more detailed explanation of this design choice in the revised version.

---

What would be the performance if reconstruction is followed by defocus deblurring in the image domain?

We performed experiments and reported results in our supplementary material (Fig. 11 and Fig. 12), where the NRKNet network was employed as SIDD network processing for defocus removal.

The results of other methods + SIDD were suboptimal. This is because:

• Defocused event streams are inherently sparse and contain degraded spatial gradients.
• Reconstructed intensity images from such events exhibit heavy artifacts and blurred textures, limiting the effectiveness of downstream image-based deblurring algorithms.
• Events capture log-intensity changes, not absolute intensity, and are affected by defocus noise, which severely impacts the quality of intermediate reconstructions.

These findings suggest that directly addressing defocus in the event domain, as EvFocus does, is more robust and efficient.

---

What is the significance of using an event camera for this task?

Thank you for your suggestion. Compared to Lou et al. [1], their work leverages a focal sweep with event cameras to reconstruct an all-in-focus image using timestamp selection and focal stack merging. This method assumes a dynamic focus adjustment and high-quality RGB input. In contrast, our setting targets scenarios where only event streams are available, especially in high-speed or low-light scenes.

Our approach reconstructs sharp images from single-pass, defocused event streams, making it suitable for settings without active focus control or RGB inputs like Lou et al. [1]. Hence, it complements Lou et al.’s method and serves different application domains.

---

We thank the reviewer again for the helpful comments and the recognition of our work. We will incorporate the suggestions and clarifications in the revised version.