Rainy Image Dataset

This repository serves as a mirror of the data used for paper “Clearing the Skies: A deep network architecture for single image rain removal” (arXiv:1609.02087) by Fu, Huang, Ding, Liao, and Paisley.

The original dataset is hosted on Baiduyun, which some users – particularly those with no working knowledge of written Chinese – might find difficult to navigate. As such, this repository serves as an alternative way to access this dataset. More information can be found at Xueyang Fu’s landing page for the paper at his personal site here: https://xueyangfu.github.io/projects/tip2017.html.

All credit for preparing this dataset goes solely to the original authors.

I additionally intend to expand on Fu et al.’s work on this dataset by providing it in a variety of pre-packaged formats, e.g. TensorFlow’s TFRecord and HDF5.

Obtaining the Dataset

Raw

If you’d like to simply retrieve the dataset files as is, there are some caveats.

This repository uses Git LFS to store the image files. As such, a direct clone will not in itself get you the image files. Attempting the clone this repository without Git LFS installed will get you files containing text pointers, rather than the image data itself.

To obtain the actual image data, first install Git LFS per instructions. Then, you have one of several options:

1. Clone the dataset repository as you would any other. This is most straightforward; however, download time will be slow due to cloning only allowing sequential file downloads.
2. Alternatively, you can achieve a faster parallel batch download by disabling Git LFS at clone time by first running:

   GIT_LFS_SKIP_SMUDGE=1 git clone https://github.com/jinnovation/rainy-image-dataset.git
       

   And then:

   git lfs pull
       

If you cloned the repository before installing Git LFS, you can download the image data using:

git lfs pull

TFRecord

Once you’ve obtained the dataset, this repository also contains a script – aptly named tfrecord.py = that will package the dataset files into a single TFRecord file to facilitate consumption into your TensorFlow model. The script interface is as follows:

Usage: tfrecord.py [OPTIONS] [INDICES]...

  Packages the dataset into a TFRecord. Arguments are the input-output
  indices to package. If no argument is provided, all indices will be
  packaged.

Options:
  --strict                 When encountering problematic image pairs, whether
                           or not to terminate.  [default: False]
  -o, --out TEXT           File name for the output .tfrecord file.  [default:
                           rain.tfrecord]
  --rainy_image_dir PATH   [default: ./rainy image/]
  --ground_truth_dir PATH  [default: ./ground truth/]
  -v, --verbosity LVL      Either CRITICAL, ERROR, WARNING, INFO or DEBUG
                           [default: DEBUG]
  --help                   Show this message and exit.

To run the script, first create a virtualenv and install dependencies listed in requirements.txt.

Note that the script is written in Python 3.

Original README text

1) This dataset contains 1,000 clean images. Each clean image was used to generate 14 rainy images with different streak orientations and magnitudes.


2) We use PhotoShop to generate rainy images: http://www.photoshopessentials.com/photo-effects/rain/


3) If this dataset helps your research, please cite our two papers that report this dataset:

  [1] X. Fu, J. Huang, D. Zeng, Y. Huang, X. Ding and J. Paisley. ¡°Removing Rain from Single Images via a Deep Detail Network¡±, CVPR, 2017.

  [2] X. Fu, J. Huang, X. Ding, Y. Liao and J. Paisley. ¡°Clearing the Skies: A deep network architecture for single-image rain removal¡±, IEEE Transactions on Image Processing, vol. 26, no. 6, pp. 2944-2956, 2017.


4) Welcome to our homepage: http://smartdsp.xmu.edu.cn/