Research map / Video / Video segmentation and tracking
A Decoupled Learning Scheme for Real-world Burst Denoising from Raw Images
Zhetong Liang, Shi Guo, Hong Gu, Huaqi Zhang, Lei Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| use two separate instances of PreP and PostP modules (decoupled) instead of shared modules in alternating training (ablation) | BDNet · Real-static | PSNR/SSIM: 44.67/0.967 -> 45.31/0.971 | Improved |
| use decoupled alternating training instead of train on Dd then fine-tune on Ds (ablation) | BDNet · Real-static | PSNR/SSIM: 45.17/0.968 -> 45.31/0.971 | No clear effect |
| use decoupled learning instead of fine-tuning on static data (perceptual quality, dynamic scenes) (ablation) | BDNet · Real-dynamic | visual quality | Improved |
| use BDNet with decoupled learning instead of KPN | burst denoising · Vimeo-200 synthetic G50 | PSNR/SSIM: 32.44/0.862 -> 34.03/0.900 | Improved |
| use BDNet instead of RIDNet | burst denoising · Vimeo-200 synthetic PG | PSNR/SSIM: 38.34/0.954 -> 39.45/0.965 | Improved |
| use BDNet instead of TOFlow | burst denoising · Vimeo-200 synthetic G25 | PSNR/SSIM: 34.99/0.902 -> 36.78/0.937 | Improved |
| increase noise level from G25 to G50 for image-domain alignment (TOFlow) | TOFlow · Vimeo-200 synthetic | PSNR: 34.99 -> 31.95 | Hurt |
| make single-frame models multi-frame with optical flow alignment and average fusion (M-UNet vs UNet) | UNet · Real-static | PSNR/SSIM: 43.85/0.954 -> 44.23/0.964 | Improved |
| make RIDNet multi-frame with optical flow alignment and average fusion (M-RIDNet vs RIDNet) | RIDNet · Real-static | PSNR/SSIM: 44.17/0.960 -> 44.54/0.968 | Improved |
| use BDNet instead of INN | burst denoising · Real-static | PSNR/SSIM: 43.95/0.964 -> 45.31/0.971 | Improved |
| train on synthetic data only (KPN) instead of adapting to real static data | KPN · Real-static | PSNR/SSIM: 39.68/0.867 | Hurt |
Closest papers by meaning
- Supervised Raw Video Denoising With a Benchmark Dataset on Dynamic Scenes (CVPR 2020 · similarity 0.89)
- Burst Denoising With Kernel Prediction Networks (CVPR 2018 · similarity 0.89)
- Deep Burst Denoising (ECCV 2018 · similarity 0.88)
- Efficient Burst Raw Denoising with Variance Stabilization and Multi-frequency Denoising Network (IJCV 2022 · similarity 0.88)
- Model-Blind Video Denoising via Frame-To-Frame Training (CVPR 2019 · similarity 0.87)
- Classic Video Denoising in a Machine Learning World: Robust, Fast, and Controllable (CVPR 2025 · similarity 0.87)
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.
