SDWNet: A Straight Dilated Network with Wavelet Transformation for Image Deblurring(offical)

1. Introduction

This repo is not only used for our paper(SDWNet) but also used for Deblur codebase. We implement a number of components that allow you to quickly implement your own model.

Paper The SDWNet has been accepted by iccvw2021, you can read the paper here.
Model

2. Folder Structure

  ---SDWNet
  |
  |- config
  |    |- model.yaml                            -> Model all traninig hyparameters with data log.
  |    |-Config.py                              -> Translate the config file to dict.
  |- data
  |    |- vanilar_dataset.py                    -> The dataset for build the LR & HR images.
  |    |- utils.py                              -> Utils for get patch and calculate the model metrics.
  |    |- augments.py                           -> Augment method for LR & HR images.
  |- model
  |    - NTIRE2021_Deblur
  |        - uniA_ELU
  |            |- layerlib_stage1               -> Model module.
  |            |- model_stage1_dual_branch_tail.py -> Main model.
  |- loss
  |   |- gendrator_loss.py                      -> Loss function define.
  |- optim
  |   |- optimizer.py                           -> Optimizer function define.
  |- train.py                                   -> Training.
  |- goprol_train.sh                             -> Training shell.
  |- inference_ddp.py                           -> Inference.
  |- inference_ddp.sh                           -> Inference shell.

3. Training

Crop the src Training LR and HR images to 480x480 by sliding window which step is 240, so we got 24 patchs form one 720x1280 images both LR and HR.
Training the model with the 416 x 416 size, use randomcrop, RGB shuffle, horizon flip, rotate and so on.
Normalize the images to Tensor with 255 but not 1. which without process the mean and std.

python -W ignore train.py \
--config_file $config_folder \
--dist-url 'tcp://127.0.0.1:8888' \
--dist-backend 'nccl' \
--multiprocessing-distributed=1 \
--world-size=1 \
--rank=0 \

4. Inference

Inference the src LR images and get the SR images

python -W ignore inference_ddp.py \
--config_file $config_folder \
--dist-url 'tcp://127.0.0.1:8989' \
--dist-backend 'nccl' \
--multiprocessing-distributed=1 \
--world-size=1 \
--rank=0 \

5. Calculate

Calculate the PSNR and SSIM

python utils/calc_psnr_ssim_official.py

If you find this repo useful for your research, please consider citing the papers

@InProceedings{
  Zou_2021_ICCV, 
  author = {Zou, Wenbin and Jiang, Mingchao and Zhang, Yunchen and Chen, Liang and Lu, Zhiyong and Wu, Yi}, 
  title = {SDWNet: A Straight Dilated Network With Wavelet Transformation for Image Deblurring}, 
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops}, 
  month = {October}, 
  year = {2021}, 
  pages = {1895-1904} 
  }

An Straight Dilated Network with Wavelet for image Deblurring

Related tags

Overview

SDWNet: A Straight Dilated Network with Wavelet Transformation for Image Deblurring(offical)

1. Introduction

2. Folder Structure

3. Training

4. Inference

5. Calculate

Owner

FlyEgle

Unofficial PyTorch implementation of Google AI's VoiceFilter system

Repository for the "Gotta Go Fast When Generating Data with Score-Based Models" paper

1st-in-MICCAI2020-CPM - Combined Radiology and Pathology Classification

DI-smartcross - Decision Intelligence Platform for Traffic Crossing Signal Control

PyTorch implementation of Octave Convolution with pre-trained Oct-ResNet and Oct-MobileNet models

Learning Optical Flow from a Few Matches (CVPR 2021)

Baseline powergrid model for NY

Synthetic Humans for Action Recognition, IJCV 2021

🥈78th place in Riiid Solution🥈

The official pytorch implementation of our paper "Is Space-Time Attention All You Need for Video Understanding?"

A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation

This is an official pytorch implementation of Lite-HRNet: A Lightweight High-Resolution Network.

Beta Shapley: a Unified and Noise-reduced Data Valuation Framework for Machine Learning

[ECCV 2020] XingGAN for Person Image Generation

Keras implementation of Real-Time Semantic Segmentation on High-Resolution Images

Automatically Build Multiple ML Models with a Single Line of Code. Created by Ram Seshadri. Collaborators Welcome. Permission Granted upon Request.

Dynamic Attentive Graph Learning for Image Restoration, ICCV2021 [PyTorch Code]

This is the repository for Learning to Generate Piano Music With Sustain Pedals

Implementation of the paper "Generating Symbolic Reasoning Problems with Transformer GANs"

for taichi voxel-challange event