Research map / Video / Video segmentation and tracking
Event-Enhanced Blurry Video Super-Resolution
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.
Introduces Ev-DeblurVSR, an event-enhanced network for blurry video super-resolution that fuses intra-frame events (for feature deblurring) and inter-frame events with optical flow (for deformable alignment).
Problem
Blurry video super-resolution (BVSR): generating sharp HR videos from blurry LR inputs, where blur removes motion information and LR frames lack high-frequency detail, causing artifacts and jitter.
Approach
Categorize events into intra-frame events (motion within exposure, used to deblur frame features via reciprocal cross-modal attention) and inter-frame events (continuous motion between frames, combined with SpyNet optical flow in a hybrid deformable alignment module), plus an edge-enhanced loss weighted by event-derived HR edge masks.
Key results
- On real NCER data, Ev-DeblurVSR surpasses FMA-Net by 2.59 dB PSNR and runs 7.28x faster (79.6 ms vs 579.8 ms per frame).
- On GoPro 4x BVSR, Ev-DeblurVSR reaches 32.51 dB PSNR / 0.9314 SSIM / 0.2041 LPIPS, exceeding EvTexture (31.00 dB) and FMA-Net (29.24 dB).
- Ablations show removing intra-frame events costs 1.19 dB and removing the i-to-e cross-modal refinement costs 0.70 dB on GoPro.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add inter-frame events to intra-frame event processing (ablation) | Ev-DeblurVSR · GoPro 4× BVSR | PSNR: 31.32 -> 32.51 | Helped |
| add intra-frame events to inter-frame event processing (ablation) | Ev-DeblurVSR · GoPro 4× BVSR | PSNR: 31.51 -> 32.51 | Helped |
| add CAB to RFD (ablation) | Ev-DeblurVSR · GoPro 4× BVSR | PSNR: 31.55 -> 32.51 | Helped |
| add cross-modal interaction to RFD (ablation) | Ev-DeblurVSR · GoPro 4× BVSR | PSNR: 31.36 -> 32.51 | Helped |
| add image-to-event refinement (ablation) | Ev-DeblurVSR · GoPro 4× BVSR | PSNR: 31.81 -> 32.51 | Helped |
| apply image-to-event refinement before event-to-image deblurring (ablation) | Ev-DeblurVSR · GoPro 4× BVSR | PSNR: 32.23 -> 32.51 | Helped |
| add event-guided alignment (ablation) | Ev-DeblurVSR · GoPro 4× BVSR | PSNR: 31.72 -> 32.51 | Helped |
| add flow-guided alignment (ablation) | Ev-DeblurVSR · GoPro 4× BVSR | PSNR: 31.53 -> 32.51 | Helped |
| add reconstruction loss (ablation) | Ev-DeblurVSR · GoPro 4× BVSR | PSNR: 32.37 -> 32.51 | Helped |
| add edge-enhanced loss (ablation) | Ev-DeblurVSR · GoPro 4× BVSR | PSNR: 32.21 -> 32.51 | Helped |
Surprising findings
- Event-based VSR methods EGVSR and EvTexture often do not beat the image-only BasicVSR++ when trained on blurry inputs, per the authors' observations.
Limitations
- Assumes frame exposure time is known and fixed; unknown or auto-exposure settings are not handled.
- Experiments limited to synthetic GoPro/BSD and one real event dataset (NCER).
Experimental setup
Datasets: GoPro BSD NCER Vid2E (simulated events)
Benchmarks: GoPro BSD NCER
Metrics: PSNR SSIM LPIPS tOF TCC #Params FLOPs runtime
Baselines: BasicVSR++ MIA-VSR IART FMA-Net eSL-Net++ EGVSR EvTexture DSTNet BSSTNet EFNet REFID
Builds on: BasicVSR++ EvTexture FMA-Net Vid2E
Closest papers by meaning
- Frequency-aware Event-based Video Deblurring for Real-World Motion Blur (CVPR 2024 · similarity 0.91)
- EvSTVSR: Event Guided Space-Time Video Super-Resolution (AAAI 2025 · similarity 0.88)
- FMA-Net: Flow-Guided Dynamic Filtering and Iterative Feature Refinement with Multi-Attention for Joint Video Super-Resolution and Deblurring (CVPR 2024 · similarity 0.87)
- Learning Scale-Aware Spatio-temporal Implicit Representation for Event-based Motion Deblurring (ICML 2024 · similarity 0.86)
- EvTexture: Event-driven Texture Enhancement for Video Super-Resolution (ICML 2024 · similarity 0.86)
- Blur-aware Spatio-temporal Sparse Transformer for Video Deblurring (CVPR 2024 · similarity 0.86)
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.