Research map / Label-efficient and robust learning / Test-time adaptation
Test-Time Fast Adaptation for Dynamic Scene Deblurring via Meta-Auxiliary Learning
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add meta-auxiliary training and five test-time reconstruction updates (ablation) | Pre-trained deblurring model · GoPro test set | PSNR: 32.30 -> 32.50 | Helped |
| add meta-auxiliary training and five test-time reconstruction updates (ablation) | Pre-trained deblurring model · HIDE test set | PSNR: 30.35 -> 30.55 | Helped |
| add self-reconstruction as an auxiliary task during joint training (ablation) | Single-scale deblurring network · GoPro and HIDE | PSNR: 31.04 -> 31.14; 28.79 -> 28.92 | Helped |
| share more weights between the deblurring and reconstruction tasks (ablation) | Single-scale deblurring network · GoPro and HIDE | PSNR: 31.07 -> 31.14; 28.86 -> 28.92 | Helped |
| use a multi-scale network instead of a single-scale network (ablation) | Deblurring network · GoPro and HIDE | PSNR: 31.04 -> 31.79; 28.79 -> 29.89 | Helped |
| add feature recurrence across scales (ablation) | Multi-scale deblurring network · GoPro and HIDE | PSNR: 31.79 -> 32.03; 29.89 -> 30.22 | Helped |
| add self-reconstruction auxiliary task (ablation) | Multi-scale network with feature recurrence · GoPro and HIDE | PSNR: 32.03 -> 32.30; 30.22 -> 30.35 | Helped |
| increase inner-loop updates and adaptation batch size (ablation) | Meta-auxiliary deblurring model · GoPro and HIDE | PSNR: 32.30; 32.38; 32.37; 32.43; 32.50; 30.37; 30.38; 30.41; 30.50; 30.55 | Helped |
| adapt only every 10 video frames instead of every frame (ablation) | Dynamic video deblurring · Adobe240 and REDS | PSNR: 32.87 -> 32.83; 29.96 -> 29.92 | Hurt |
| naively update the pre-trained model using the auxiliary loss without meta-auxiliary training (ablation) | Pre-trained deblurring model · GoPro, HIDE, Adobe240, and REDS | PSNR | Hurt |
Closest papers by meaning
- Clean Images are Hard to Reblur: Exploiting the Ill-Posed Inverse Task for Dynamic Scene Deblurring (ICLR 2022 · similarity 0.88)
- CoA: Towards Real Image Dehazing via Compression-and-Adaptation (CVPR 2025 · similarity 0.85)
- DeblurDiff: Real-Word Image Deblurring with Generative Diffusion Models (NeurIPS 2025 · similarity 0.85)
- Self-Bootstrapping for Versatile Test-Time Adaptation (ICML 2025 · similarity 0.84)
- Fast Adaptation for Human Pose Estimation via Meta-Optimization (CVPR 2024 · similarity 0.84)
- ID-Blau: Image Deblurring by Implicit Diffusion-based reBLurring AUgmentation (CVPR 2024 · 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.