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

Code for Neural-GIF: Neural Generalized Implicit Functions for Animating People in Clothing(ICCV21)

TensorFlow-LiveLessons - "Deep Learning with TensorFlow" LiveLessons

The official repository for "Score Transformer: Generating Musical Scores from Note-level Representation" (MMAsia '21)

Pytorch implementation of Straight Sampling Network For Point Cloud Learning (ICIP2021).

A Physics-based Noise Formation Model for Extreme Low-light Raw Denoising (CVPR 2020 Oral & TPAMI 2021)

Code and real data for the paper "Counterfactual Temporal Point Processes", available at arXiv.

Human4D Dataset tools for processing and visualization

This project aims at providing a concise, easy-to-use, modifiable reference implementation for semantic segmentation models using PyTorch.

Official repository of ICCV21 paper "Viewpoint Invariant Dense Matching for Visual Geolocalization"

Official Pytorch implementation of 'GOCor: Bringing Globally Optimized Correspondence Volumes into Your Neural Network' (NeurIPS 2020)

A-ESRGAN aims to provide better super-resolution images by using multi-scale attention U-net discriminators.

PyTorch implementation of paper: HPNet: Deep Primitive Segmentation Using Hybrid Representations.

This PyTorch package implements MoEBERT: from BERT to Mixture-of-Experts via Importance-Guided Adaptation (NAACL 2022).

Patch-Diffusion Code (AAAI2022)

CyTran: Cycle-Consistent Transformers for Non-Contrast to Contrast CT Translation

SymmetryNet: Learning to Predict Reflectional and Rotational Symmetries of 3D Shapes from Single-View RGB-D Images

This repository contains tutorials for the py4DSTEM Python package

A small demonstration of using WebDataset with ImageNet and PyTorch Lightning

Development Kit for the SoccerNet Challenge

Non-Vacuous Generalisation Bounds for Shallow Neural Networks