A state-of-the-art, simple and fast network for Deep Video Denoising which uses no motion compensation.

Oral presentation at CVPR2020. CVPR publication page

Previous deep video denoising algorithm: DVDnet

This source code provides a PyTorch implementation of the FastDVDnet video denoising algorithm, as in Tassano, Matias and Delon, Julie and Veit, Thomas. "FastDVDnet: Towards Real-Time Deep Video Denoising Without Flow Estimation", arXiv preprint arXiv:1907.01361 (2019).

You can download several denoised sequences with our algorithm and other methods here

The 2017 DAVIS dataset was used for training. You can find a list with the names of the 480p sequences employed here. The dataloader needs the sequences in mp4 format. You can find the converted .mp4 files under the training folder here.

Note: when converting the sequences one has to pay particular attention to the 'crf' and 'keyint' ffmpeg parameters to avoid strong compression. For the code to convert the image sequences see this gist

Two testsets are used in the paper: Set8 and the 2017 DAVIS testset.

Set8 is composed of 8 sequences: 4 sequences from the Derf 480p testset ("tractor", "touchdown", "park_joy", "sunflower") plus other 4 540p sequences. You can find these under the test_sequences folder here.

FastDVDnet is orders of magnitude faster than other state-of-the-art methods

Left: Input noise sigma 40 denoised with FastDVDnet (sorry about the dithering due to gif compression)

Right: PSNRs on the DAVIS testset, Gaussian noise and clipped Gaussian noise

You can use this Colab notebook to replicate the results

The code runs on Python +3.6. You can create a conda environment with all the dependecies by running (Thanks to Antoine Monod for the .yml file)

conda env create -f requirements.yml -n <env_name>

Note: this project needs the NVIDIA DALI package for training. The tested version of DALI is 0.10.0. If you prefer to install it yourself (supposing you have CUDA 10.0), you need to run

pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/cuda/10.0 nvidia-dali==0.10.0 

If you want to denoise an image sequence using the pretrained model you can execute

test_fastdvdnet.py \
	--test_path <path_to_input_sequence> \
	--noise_sigma 30 \
	--save_path results

NOTES

• The image sequence should be stored under <path_to_input_sequence>
• The model has been trained for values of noise in [5, 55]
• run with --no_gpu to run on CPU instead of GPU
• run with --save_noisy to save noisy frames
• set max_num_fr_per_seq to set the max number of frames to load per sequence
• to denoise clipped AWGN run with --model_file model_clipped_noise.pth
• run with --help to see details on all input parameters

If you want to train your own models you can execute

train_fastdvdnet.py \
	--trainset_dir <path_to_input_mp4s> \
	--valset_dir <path_to_val_sequences> \
	--log_dir logs

NOTES

• As the dataloader in based on the DALI library, the training sequences must be provided as mp4 files, all under <path_to_input_mp4s>
• The validation sequences must be stored as image sequences in individual folders under <path_to_val_sequences>
• run with --help to see details on all input parameters

Copying and distribution of this file, with or without modification, are permitted in any medium without royalty provided the copyright notice and this notice are preserved. This file is offered as-is, without any warranty.

• Author : Matias Tassano mtassano at gopro dot com
• Copyright : (C) 2019 Matias Tassano
• Licence : GPL v3+, see GPLv3.txt

The sequences are Copyright GoPro 2018