Efficient Multi-Stage Video Denoising With Recurrent Spatio-Temporal Fusion.

EMVD is an efficient video denoising method which recursively exploit the spatio temporal correlation inherently present in natural videos through multiple cascading processing stages applied in a recurrent fashion, namely temporal fusion, spatial denoising, and spatio-temporal refinement.

This repo. is an unofficial version od EMVD mentioned by Matteo Maggioni, Yibin Huang, Cheng Li, Shuai Xiao, Zhongqian Fu, Fenglong Song in CVPR 2021.

It is a Pytorch implementation.

• https://openaccess.thecvf.com/content/CVPR2021/papers/Maggioni_Efficient_Multi-Stage_Video_Denoising_With_Recurrent_Spatio-Temporal_Fusion_CVPR_2021_paper.pdf
• https://openaccess.thecvf.com/content/CVPR2021/supplemental/Maggioni_Efficient_Multi-Stage_Video_CVPR_2021_supplemental.pdf

1. PyTorch>=1.6
2. Numpy
3. scikti-image
4. tensorboardX (for visualization of loss, PSNR and images)
5. torchstat （for computing GFLOPs）

1. config.py is the code for setting hyperparameters.
2. dataset.py and load_data.py is the code for loading data from dataset.
3. train.py is the code for training process
4. inference.py is the code for validation process.
5. models.py and ./isp/ISP_CNN.pth is called by inference.py for converting .tiff to .png, which refer to the code RViDeNet(https://github.com/cao-cong/RViDeNet).

CRVD Dataset (https://github.com/cao-cong/RViDeNet)

modify data_root in config.py, and gt_name/noisy_name in function decode_data inload_data.py, and run train.py for training process. After convergence， run inference.py for validation process.

ISO average raw psnr:42.02, iso frame average raw ssim:0.9800 in CRVD datasets (~5.38GFLPs), which is still lower than the experiment results mentioned in paper.

This implementations are inspired by following projects:

• [RViDeNet] (https://github.com/cao-cong/RViDeNet)

Many thanks for coming here! It will be highly appreciated if you offer any suggestion.

Support me by starring or forking this repo., please.