yujilin/raindirection-and-real3000-dataset

This repository contains an introduction and links to RainDirection and Real3000 datasets for paper Unpaired Learning for Deep Image Deraining with Rain Direction Regularizer, Authors:Yang Liu, Ziyu Yue ,Jinshan Pan, Zhixun Su, Dalian University of Technology and Nanjing University of Science and Technology

RainDirection-and-Real3000-Dataset

This repository contains an introduction and links to RainDirection and Real3000 datasets for paper "Unpaired Learning for Deep Image Deraining with Rain Direction Regularizer". Authors: Yang Liu, Ziyu Yue ,Jinshan Pan, Zhixun Su, Dalian University of Technology and Nanjing University of Science and Technology. In the part about dataset of this work, I wrote a program to generate rainy images and rain streak maps with rain streak direction information in batches by imitating the process of PS rain making, and output their corresponding label files of rain streak direction information at the same time. And I collected 3,000 real world rain images with no ground truth from pictures and videos on the Internet.This work was published in the IEEE International Conference on Computer Vision 2021(ICCV2021). You can get the paper from here.

Data overview

We have maked a new rainy dataset RainDirection with the direction information of rain streak and a new real world rainy dataset Real3000. You can download RainDirection and Real3000 dataset from this BaiduYun link (Extract code: 33e8).

RainDirection

The RainDirection dataset is a synthetic rainy dataset. The rainy images in RainDirection are obtained by adding clean images from Flick2K and DIV2K dataset with synthetic labeled rain maps according to the rain model O(x) = B(x) + R(x), and the rain streak map is obtained by the direction fuzzy kernel acting on the sparse Gaussian noise map. Each rainy image is assigned with a direction label, the horizontal Angle of the image is zero to the right and increases the angle counterclockwise. These direction labels are used to calculate the direction loss during training. The training and test set of RainDirection contains 2920 and 430 images, respectively.

The folder structure is as follows:

rainy_direction_dataset
  RainDirection2920train
      direction
          direction.txt
      latent
          1.png
          2.png
          ...
          ...
          2920.png
      rainy
          1.png
          2.png
          ...
          ...
          2920.png
      streak
          1.png
          2.png
          ...
          ...
          2920.png

  RainDirection430test
      latent
          1.png
          2.png
          ...
          ...
          430.png
      rainy
          1.png
          2.png
          ...
          ...
          430.png
      streak
          1.png
          2.png
          ...
          ...
          430.png

Real3000

The Real3000 dataset contains 3000 real rainy images without ground truth images collected from the internet and captured by a Canon EOS 6D camera. The training and test set contains 2700 and 300 diverse natural outdoor images, respectively. Images numbered 1705 to 2004 were selected as the test set.

The folder structure is as follows:

real3000
  1.png
  2.png
  ...
  ...
  3000.png

Citation:

@inproceedings{liu2021unpaired,
  title={Unpaired Learning for Deep Image Deraining With Rain Direction Regularizer},
  author={Liu, Yang and Yue, Ziyu and Pan, Jinshan and Su, Zhixun},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={4753--4761},
  year={2021}
}

Contact:

If you have any question, please email [email protected]