ReMoNet: Recurrent Multi-output Network for Efficient Video Denoising

Implementation of
"ReMoNet: Recurrent Multi-output Network for Efficient Video Denoising"
Liuyu Xiang, Jundong Zhou, Jirui Liu, Zerun Wang, Haidong Huang, Jie Hu, Jungong Han, Yuchen Guo, Guiguang Ding;

in AAAI 2022

Requirements

• PyTorch >=1.1.0
• scikit-learn
• opencv

Data Preparation

Follow FastDVDNet for data preparation.

Experiments with Gaussian Noise

Step 1: Change trainset_dir & valset_dir to your path

Step 2: python train_ReMoNet.py --log_dir path/to/logdir

Experiments with ISP simulated Noise

Step 1: First use ISP_sim_code to generate ISP simulated data

Step 1.1 Change im_folder and out_dir in generate_dataset_patches.py

Step 1.2 python generate_dataset_patches.py

The result is saved in args.out_dir/train.h5

Note: The process could be slow and take up to ~12 hours to finish for DAVIS-train. Once finished, the training is fast ^_^.

Step 2: python train_ReMoNet.py --log_dir path/to/logdir

Citation

If you find our work useful for your research, please consider citing the following paper:

@article{article,
author = {Xiang, Liuyu and Zhou, Jundong and Liu, Jirui and Wang, Zerun and Huang, Haidong and Hu, Jie and Han, Jungong and Guo, Yuchen and Ding, Guiguang},
year = {2022},
month = {06},
pages = {2786-2794},
title = {ReMoNet: Recurrent Multi-Output Network for Efficient Video Denoising},
volume = {36},
journal = {Proceedings of the AAAI Conference on Artificial Intelligence},
doi = {10.1609/aaai.v36i3.20182}
}

Contact

If you have any questions, please feel free to contact xiangly55@foxmail.com

Acknowledgement

The code is partly based on FastDVDNet. The ISP simulated pipeline is borrowed from camera_sim. Many thanks to them!