Skip to content

ECCV2020 - Practical Deep Raw Image Denoising on Mobile Devices

License

Notifications You must be signed in to change notification settings

MegEngine/PMRID

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Practical Mobile Raw Image Denoising (PMRID)

Code and dataset for ECCV20 paper Practical Deep Raw Image Denoising on Mobile Devices.

Dataset

Downloads

Usage

The dataset includes two 7zip files:

  • reno10x_noise.7z contains DNG raw images shot by an OPPO Reno 10x phone for noise parameter estimation (refer Sec 3.1 and 5.1 in the paper)
  • PMRID.7z is the benchmark dataset described in Sec 5.2 in the paper

The structure of PMRID.7z is

- benchmark.json  # meta info
- Scene1/
  \- Bright/
     \- exposure-case1/ 
         \- input.raw   # RAW data for noisy image in uint16
          - gt.raw      # RAW data for clean image in uint16
      + case2/
  + Dark/
+ Secne2/

All metadata for images are listed in benchmark.json:

{
   "input": "path/to/noisy_input.raw",
   "gt": "path/to/clean_gt.raw",
   "meta": {
       "name": "case_name",
       "scene_id": "scene_name",
       "light": "light condition",
       "ISO": "ISO",
       "exp_time": "exposure time",
       "bayer_pattern": "BGGR",
       "shape": [3000, 4000],
       "wb_gain": [r_gain, g_gain, b_gain],
       "CCM": [   # 3x3 color correction matrix
           [c11, c12, c13], 
           [c21, c22, c23], 
           [c31, c32, c33]
       ],
       "ROIs": [  # patch ROIs to calculate PSNR and SSIM, x0 is topleft
           [topleft_w, topleft_h, bottomright_w, bottomright_h]
       ]
   }
}

Pre-trained Models and Benchmark Script

Both PyTorch and MegEngine pre-trained models are provided in the models directory. The benchmark script is written for models trained with MegEngine. Python >= 3.6 is required to run the benchmark script.

pip install -r requirements.txt
python3 run_benchmark.py --benchmark /path/to/PMRID/benchmark.json models/mge_pretrained.ckp

Citation

@inproceedings{wang2020,
	title={Practical Deep Raw Image Denoising on Mobile Devices},
	author={Wang, Yuzhi and Huang, Haibin and Xu, Qin and Liu, Jiaming and Liu, Yiqun and Wang, Jue},
	booktitle={European Conference on Computer Vision (ECCV)},
	year={2020},
	pages={1--16}
}

About

ECCV2020 - Practical Deep Raw Image Denoising on Mobile Devices

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages