Research map / 3D vision / Depth and stereo
EvFocus: Learning to Reconstruct Sharp Images from Out-of-Focus Event Streams
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes EvFocus, a network that reconstructs sharp images from defocused event streams using a ConvLSTM temporal encoder, a blur-aware two-branch decoder, and a re-defocus/reconstruction module, trained on a new simulated defocus event dataset.
Problem
Event cameras produce sparse event streams that defocus blur degrades (fewer and weaker events), and existing event-based reconstruction and deblurring methods do not handle defocus blur effectively.
Approach
An event-based out-of-focus camera model and simulator generate defocus event streams; EvFocus separates blur features from alignment features with a two-branch decoder and adds a self-supervised re-defocus branch so the network learns the defocus blur distribution.
Key results
- On simulated defocus data, EvFocus achieves mean MSE 0.027, SSIM 0.468, LPIPS 0.166, outperforming ET-Net* (0.028, 0.396, 0.175).
- On real-world DAVIS 346 data, EvFocus gets BRISQUE 27.136 and NIQE 6.345, the best among compared methods (best baseline BRISQUE 31.855 for ET-Net*).
- Ablation shows removing the L1 re-defocus supervision drops SSIM from 0.604 to 0.403.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EvFocus instead of the strongest baseline, E2VID+* | Event reconstruction on simulated event data · Simulated event data, mean | SSIM: 0.433 -> 0.468 | Helped |
| Use EvFocus instead of the strongest baseline, ET-Net* | Event reconstruction on simulated event data · Simulated event data, mean | MSE: 0.028 -> 0.027 | Helped |
| Use EvFocus instead of the strongest baseline, ET-Net* | Event reconstruction on simulated event data · Simulated event data, mean | LPIPS: 0.175 -> 0.166 | Helped |
| Use EvFocus instead of ET-Net* | Event reconstruction on real-world data · Real-world data | BRISQUE: 31.855 -> 27.136 | Helped |
| Use EvFocus instead of ET-Net* | Event reconstruction on real-world data · Real-world data | NIQE: 6.868 -> 6.345 | Helped |
| Add Cross-Modal Fusion (CF) (ablation) | EvFocus · Simulated event data | SSIM: 0.603 -> 0.604 | No clear effect |
| Add sharp-ground-truth supervision for Falign (ablation) | EvFocus · Simulated event data | MSE: 0.024 -> 0.021 | Helped |
| Add L1 re-defocus supervision (ablation) | EvFocus · Simulated event data | SSIM: 0.403 -> 0.604 | Helped |
| Add Reconstruction Module (ablation) | EvFocus · Simulated event data | SSIM: 0.601 -> 0.604 | Helped |
Surprising findings
- Removing the L1 re-defocus supervision sharply lowers SSIM (0.604 to 0.403) while MSE and LPIPS barely change.
Limitations
- Reconstruction quality degrades for blur kernels larger than size 7 according to the authors' own analysis.
- Real-world evaluation uses only no-reference metrics (BRISQUE, NIQE) on 7 scenes.
Experimental setup
Datasets: simulated defocus event streams generated with ESIM real-world DAVIS 346 defocus event recordings
Benchmarks: simulated defocus event data (5 blur levels) real-world DAVIS 346 defocus event data
Metrics: SSIM MSE LPIPS BRISQUE NIQE
Baselines: E2VID E2VID+ ET-Net SPADE-E2VID MIMO-UNet
Builds on: E2VID ConvLSTM ESIM Restormer-style transformer (MHCA)
Closest papers by meaning
- Event Fields: Capturing Light Fields at High Speed, Resolution, and Dynamic Range (CVPR 2025 · similarity 0.86)
- One-Step Event-Driven High-Speed Autofocus (CVPR 2025 · similarity 0.86)
- Towards Robust Event-guided Low-Light Image Enhancement: A Large-Scale Real-World Event-Image Dataset and Novel Approach (CVPR 2024 · similarity 0.86)
- EventSplat: 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering (CVPR 2025 · similarity 0.85)
- Mitigating Motion Blur in Neural Radiance Fields with Events and Frames (CVPR 2024 · similarity 0.85)
- Event-3DGS: Event-based 3D Reconstruction Using 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.