Research map / Video / Video segmentation and tracking
Frequency-aware Event-based Video Deblurring for Real-World Motion Blur
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 a frequency-aware event-based video deblurring framework with cross-modal frequency fusion and event-guided temporal alignment, plus a new real-world event-video deblurring dataset (REVD).
Problem
Video deblurring under severe motion blur loses motion information during exposure; existing event-based deblurring largely handles single frames and does not exploit long-range temporal dependencies in video.
Approach
Fuse event and blur-frame features in the frequency domain (spatial and global channel frequency filtering) and align video frames using event-guided local-windowed propagation and bidirectional long-range temporal fusion, without optical flow.
Key results
- On GoPro the method reaches 36.70 dB PSNR / 0.978 SSIM, above the best event-based baseline EFNet† at 35.87 dB.
- On the real-world REVD dataset it reaches 32.99 dB PSNR, 1.02 dB above the second-best UEVD (31.97 dB).
- Adding the FCFE module to the baseline improves PSNR by 2.25 dB in the GoPro ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed frequency-aware event-based video deblurring method instead of EFNet† | Event-based video deblurring · GoPro | PSNR: 35.87 -> 36.70 | Helped |
| Use the proposed frequency-aware event-based video deblurring method instead of UEVD | Event-based video deblurring · REVD | PSNR: 31.97 -> 32.99 | Helped |
| Add FCFE (ablation) | GoPro ablation baseline · GoPro | PSNR: 31.14 -> 33.39 | Helped |
| Add ELTP (ablation) | GoPro ablation baseline · GoPro | PSNR: 31.14 -> 35.36 | Helped |
| Add BTFF (ablation) | GoPro ablation baseline · GoPro | PSNR: 31.14 -> 35.59 | Helped |
| Add FCFE and ELTP (ablation) | GoPro ablation baseline · GoPro | PSNR: 31.14 -> 36.08 | Helped |
| Add FCFE and BTFF (ablation) | GoPro ablation baseline · GoPro | PSNR: 31.14 -> 36.45 | Helped |
| Add ELTP and BTFF (ablation) | GoPro ablation baseline · GoPro | PSNR: 31.14 -> 36.60 | Helped |
| Add FCFE, ELTP, and BTFF (ablation) | GoPro ablation baseline · GoPro | PSNR: 31.14 -> 36.70 | Helped |
Surprising findings
- The frame-based RNN-MBP reaches 31.77 dB on REVD, comparable to event-based methods, which the authors attribute to severe event noise and saturation in real data.
Limitations
- Real-world event noise and saturated pixels limit restoration quality on real data.
- Training and test scale of the REVD dataset is modest (21 sequences).
Experimental setup
Datasets: GoPro REVD ESIM-simulated events
Benchmarks: GoPro REVD
Metrics: PSNR SSIM inference time (ms) FLOPs
Baselines: IFRNN ESTRNN CDVD-TSP MMP-RNN STDAN ERDN DSTNet eSL-Net REDNet D2Nets UEVD EFNet
Builds on: EFNet UEVD RNN-MBP
Closest papers by meaning
- Event-Enhanced Blurry Video Super-Resolution (AAAI 2025 · similarity 0.91)
- Learning Scale-Aware Spatio-temporal Implicit Representation for Event-based Motion Deblurring (ICML 2024 · similarity 0.90)
- Motion Deblurring via Spatial-Temporal Collaboration of Frames and Events (AAAI 2024 · similarity 0.89)
- Diffusion-based Event Generation for High-Quality Image Deblurring (CVPR 2025 · similarity 0.88)
- Motion-adaptive Transformer for Event-based Image Deblurring (AAAI 2025 · similarity 0.88)
- Latency Correction for Event-guided Deblurring and Frame Interpolation (CVPR 2024 · similarity 0.88)
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.