Research map / Video / Video segmentation and tracking
AverNet: All-in-one Video Restoration for Time-varying Unknown Degradations
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 AverNet, a recurrent all-in-one video restoration network with prompt-guided alignment and prompt-conditioned enhancement to handle time-varying unknown degradations.
Problem
Video restoration when multiple unknown degradation types and levels vary over time within a single video (time-varying unknown degradations, TUD).
Approach
Input-conditioned prompts guide deformable-convolution alignment (PGA) to handle pixel shifts from time-varying degradations, and key-frame prompts condition feature enhancement (PCE) within a bidirectional recurrent propagation scheme, turning multi-degradation restoration into a conditional problem.
Key results
- On DAVIS-test with t=6, AverNet reaches 34.07 dB PSNR / 0.9333 SSIM, versus 33.99 dB for RVRT and 32.46 dB for AirNet.
- Ablation on DAVIS-test: removing PCE drops PSNR by 1.10 dB and removing PGA by 1.50 dB relative to the full model (34.09 dB).
- Under noise&blur on DAVIS-test, AverNet outperforms RVRT by 0.34 dB PSNR while running in less than half the time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use AverNet instead of RVRT | video restoration on DAVIS-test, t=6/12/24 · DAVIS-test, t=6/12/24 | PSNR: 33.99 -> 34.07; 33.98 -> 34.09; 34.10 -> 34.28 | Helped |
| use AverNet instead of RVRT | video restoration on Set8, t=6/12/24 · Set8, t=6/12/24 | PSNR: 31.73 vs 31.73; 31.39 -> 31.47; 32.47 -> 32.45 | Mixed |
| add PGA to the method (ablation) | AverNet on DAVIS-test and Set8, t=12 · DAVIS-test and Set8 | PSNR: 32.59 -> 34.09; 30.14 -> 31.47 | Helped |
| add PCE to the method (ablation) | AverNet on DAVIS-test and Set8, t=12 · DAVIS-test and Set8 | PSNR: 32.99 -> 34.09; 29.80 -> 31.47 | Helped |
| add both PGA and PCE prompt guidance/conditioning (ablation) | AverNet on DAVIS-test and Set8, t=12 · DAVIS-test and Set8 | PSNR: 32.43 -> 34.09; 27.99 -> 31.47 | Helped |
| increase PCE key-frame interval from T=6 to T=12 or T=24 (ablation) | AverNet · DAVIS-test | PSNR: 34.09; 34.15; 34.14 | No clear effect |
| increase PCE key-frame interval from T=6 to T=12 or T=24 (ablation) | AverNet · Set8 | PSNR: 31.47 -> 31.02; 31.47 -> 31.02 | Hurt |
Surprising findings
- Several all-in-one image restoration methods trained per frame (e.g., AirNet, PromptIR) are competitive, but video-specific recurrent methods trained from scratch can underperform on some sets, e.g., EDVR scores far lower on DAVIS-test.
Limitations
- Only three degradation families (noise, blur, compression) studied; rain, haze and snow not covered.
- Training relies on synthesized degradations, which may not capture the complexity of real-world corruption.
- No error bars reported due to computational cost.
Experimental setup
Datasets: DAVIS Set8
Benchmarks: DAVIS-test Set8
Metrics: PSNR SSIM runtime parameters
Baselines: WDiffusion TransWeather AirNet PromptIR EDVR BasicVSR++ Shift-Net RVRT
Builds on: BasicVSR++ PromptIR EDVR SPyNet
Closest papers by meaning
- Test-Time Degradation Adaptation for Open-Set Image Restoration (ICML 2024 · similarity 0.85)
- Degradation-Aware Feature Perturbation for All-in-One Image Restoration (CVPR 2025 · similarity 0.85)
- Decoupling Degradations with Recurrent Network for Video Restoration in Under-Display Camera (AAAI 2024 · similarity 0.84)
- Visual-Instructed Degradation Diffusion for All-in-One Image Restoration (CVPR 2025 · similarity 0.84)
- UP-Restorer: When Unrolling Meets Prompts for Unified Image Restoration (AAAI 2025 · similarity 0.84)
- Learning Where and What to Restore for Composite Image Restoration (NeurIPS 2026 · similarity 0.84)
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.