A decoupled learning scheme for training a burst denoising network for real-world denoising.

[Paper][Supp]

Note: we only provides the code for real-world noise removal.

• Pytorch == 1.2.0
• Cuda == 10.1
• Python == 3.7

• Dynamic video dataset. This dataset corresponds to the D_d in the main paper which provides dynamic contents for alignment learnig.
• Real-world static burst dataset. This dataset corresponds to the D_s in the main paper which provides real-world noise for learning.

• Real-static.
• Real-dynamic (to be finished)

link

This project is based on the EDVR project from https://github.com/xinntao/EDVR/tree/old_version. For details, you may refer to that project. We will simplify the code in the future.

1. Download the training datasets, including Dynamic video dataset and Real-world static burst dataset, and unzip them to the path datasets.
2. Download the testing sets. Currently we only have Real-static and we will update Real-dynamic set in the future. Download them and unzip them to the path datasets.
3. For training, python train_BDNet.py
4. For testing, you may download the pretrained models first and put it in pretrained_model folder. Then python test_BDNet.py

Zhetong Liang zhetong.liang@connect.polyu.hk

@inproceedings{liang2020bdnet, title={A Decoupled Learning Scheme for Real-worldBurst Denoising from Raw Images}, author={Zhetong Liang, Shi Guo, Hong Gu, Huaqi Zhang, and Lei Zhang}, booktitle={ECCV}, year={2020} }