TODO LIST:
- Train the network, and upload the trained model
- write the documents
- write the code of single-stage model, single-scale model to verify the superiority of FastDVDNet
This repo. is an unofficial version of FastDVDNet:ToWards Real-Time Video Denoising Without Explicit Motion Estimation by making use of PyTorch.
The dataset of Vimeo-90K is employed for training, whose size is about 82G. The dataset consists of about 90K videos and all of them include 7 frames with large motion.
You can choose the dataset you like to train or validate the effectiveness of FastDVDNet, such as the dataset indicated in the paper (I don't want to spend more time downloading it so I choose the Vimeo-90K.).
- PyTorch>=1.0.0
- Numpy
- scikti-image
- tensorboardX (for visualization of loss, PSNR and SSIM)
- Run
pip install -r requirements.txt
firstly to install the packages. data_provider.py
is the code for loading data from dataset. You can modify the code to adapt your dataset if it is not Vimeo-90K.train_eval.py
is the code for train and validation process. The validation process is activated by--eval
, otherwise, it is work on the train mode.- Script for restarting the model can be follow,
CUDA_VISIBLE_DEVICES=1,2 python train_eval.py --cuda -nw 16 --frames 5 -s 96 -bs 64 -lr 1e-4 --restart
- As for validation mode,
CUDA_VISIBLE_DEVICES=1,2 python train_eval.py --cuda --eval
- FastDVDNet:ToWards Real-Time Video Denoising Without Explicit Motion Estimation
- DVDNet: A Fast Network For Deep Video Denoising
- Code of U-Net is based on my previous work
Support me by starring or forking this repo., please.