rainy-image-dataset

rainy-image-dataset copied to clipboard

#+TITLE: Rainy Image Dataset #+AUTHOR: Jonathan Jin

This repository serves as a mirror of the data used for paper [[https://arxiv.org/abs/1609.02087]["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 [[https://www.tensorflow.org/guide/datasets#reading_from_files][TFRecord]] and [[https://www.hdfgroup.org/solutions/hdf5/][HDF5]].

• Obtaining the Dataset

** Raw

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

This repository uses [[https://git-lfs.github.com/][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 [[https://git-lfs.github.com/][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 [[https://github.com/git-lfs/git-lfs/wiki/Tutorial#pulling-and-cloning][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: #+begin_src bash :eval never GIT_LFS_SKIP_SMUDGE=1 git clone https://github.com/jinnovation/rainy-image-dataset.git #+end_src

   And then: #+begin_src bash :eval never git lfs pull #+end_src

If you cloned the repository before installing Git LFS, you can download the image data using: #+begin_src bash :eval never git lfs pull #+end_src

** 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 [[https://www.tensorflow.org/guide/datasets#reading_from_files][TFRecord]] file to facilitate consumption into your TensorFlow model. The script interface is as follows:

#+begin_src 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. #+end_src

To run the script, first [[https://virtualenv.pypa.io/en/stable/][create a virtualenv]] and install dependencies listed in =requirements.txt=.

Note that the script is written in Python 3.

• Original README text

  #+begin_src 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/ #+end_src