CDR Converter

We provide the python script to convert the dataset proposed in CDR: A Categorized and Diverse Real-World Reflection Removal Dataset based on different options.

Setup

The script requires Python 3.5+ and cv2. Please download our data (~7.45GB) from this link. We have 1063 triplets (M, R, T) in total.

Usage

python convert.py --datapath <DATAPATH> --csvpath <CSVPATH> --output <OUTPUTDIR> --[options]

You can choose to generate one (or a subset) of our dataset by setting the following self-explanatory arguments:

• --train
• --val
• --test
• --type
• --reflection
• --ghost
• --motion

For example, if you want to down only train set,

python convert.py --datapath <DATAPATH> --csvpath <CSVPATH> --output <OUTPUTDIR> --train

Note that these arguments can be combined to generate a set satisfying all options,

python convert.py --datapath <DATAPATH> --csvpath <CSVPATH> --output <OUTPUTDIR> --test --type SRST --reflection medium

will generate testset with SRST type AND medium reflection.

Considering some methods may require input image in size of a multiple 32, we also provide an argument --crop32, which will generate images in size of its nearest 32's multiples.

Some crops in our dataset may have large ratio of longer side / shorter side, you can remove those crops by --remove_extreme

Our dataset is in high resolution, so we also support downsampling option by specifying --downsample_scale argument followed by an integer.

Output

You must specify the output folder with --output argument.

Folders structure

You should expect the original data structure looks like

data/
└── isprgb_crop
    └── with_gt
        ├── C1
        ├── C10
        ├── C11
        ├── C2
        ├── C3
        ├── C4
        ├── C5
        ├── C6
        ├── C7
        ├── C8
        ├── C9
        ├── H1
        ├── N1
        ├── N2
        ├── N3
        ├── N4
        ├── N5
        ├── N6
        └── N7

• isprgb_crop: cropped data inside valid regions

Each leaf directory will contain .png files accordingly. Also, there is a four unique digit number (e.g. 5532, 5531) for each M and R image, while the corresponding T image is named as "M_R" (e.g. 5532_5531).

For normal benchmarking (as written in our script), only isprgb_crop/ folder will be used, so we only make this folder public for the first release.

Citation

If you find this dataset or code useful, please kindly reference: