Research map / Image generation / Image restoration
Joint Demosaicking and Denoising in the Wild: The Case of Training Under Ground Truth Uncertainty
Jierun Chen, Song Wen, S.-H. Gary Chan
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 |
|---|---|---|---|
| Train with Wild-JDD instead of the strongest reported baseline | Joint demosaicking and denoising models · Kodak, McMaster, WED-CDM, MIT moire, Urban100 (σ=10) | PSNR: 33.32 -> 33.81; 33.18 -> 33.53; 32.76 -> 33.44; 30.09 -> 30.30; 32.14 -> 32.42 | Improved |
| Train with Wild-JDD rather than the strongest baseline | Joint demosaicking and denoising models · MSR realistic raw, linear | PSNR: 40.07 -> 40.16 | Improved |
| Train with Wild-JDD rather than the strongest baseline | Joint demosaicking and denoising models · MSR realistic raw, sRGB | PSNR: 34.24 -> 34.34 | Improved |
| Use ELBO loss instead of MSE loss (ablation) | Wild-JDD · Kodak, McMaster, WED-CDM (σ=10) | PSNR: 33.77 -> 33.81; 33.45 -> 33.53; 33.37 -> 33.44 | Improved |
| Apply self-ensemble with flipping and rotation at test time (ablation) | Wild-JDD · Synthetic benchmarks and MSR realistic raw | PSNR: 33.81 -> 33.88 (Kodak, σ=10); 40.16 -> 40.36 (linear raw); 34.34 -> 34.59 (sRGB raw) | Improved |
| Fine-tune on out-of-distribution input using it as a weakly informative prior (ablation) | Wild-JDD · McM dataset, Uniform, Poisson-Gaussian, and Brown-Gaussian noise | PSNR: 0.1∼0.3 dB | Improved |
| Continue fine-tuning for too many iterations (ablation) | Wild-JDD fine-tuning · McM dataset, out-of-distribution noise | PSNR | Hurt |
| Use the confidence-interval mask during fine-tuning (ablation) | Wild-JDD fine-tuning · Visual edge preservation | Edge sharpness | Improved |
| Use Wild-JDD rather than DMCNN-VD-Tr | Joint demosaicking and denoising models · MSR realistic raw, linear | SSIM: 0.981 -> 0.980 | Hurt |
| Use Wild-JDD rather than SEM | Joint demosaicking and denoising models · MSR realistic raw, sRGB | SSIM: 0.960 -> 0.945 | Hurt |
Closest papers by meaning
- Not All Noise Is Harmful: Towards Perception Aware and Controllable RAW Image Joint Denoising and Demosaicing (NeurIPS 2026 · similarity 0.91)
- Joint Demosaicking and Denoising by Fine-Tuning of Bursts of Raw Images (ICCV 2019 · similarity 0.91)
- Deep Image Demosaicking using a Cascade of Convolutional Residual Denoising Networks (ECCV 2018 · similarity 0.90)
- Joint Demosaicing and Denoising With Self Guidance (CVPR 2020 · similarity 0.89)
- NERDS: A General Framework to Train Camera Denoisers from Raw-RGB Noisy Image Pairs (ICLR 2023 · similarity 0.88)
- Toward Convolutional Blind Denoising of Real Photographs (CVPR 2019 · 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.
