This tool generates an augmented, polygon-based, dataset from the Form Understanding in Noisy Scanned Documents dataset.
The resulting dataset will contain the same annotations, but the images will be augmented with a set of transformations such as perspective, brightness, contrast, etc. Plus adding random lines and "reflections".
The idea is to generate a dataset that mimics real world images taken by users with their mobile phones, instead of an scanned document dataset. For that, this dataset's annotations are polygon-based, instead of the usual bounding boxes. This means that instead of having two coordinates, we have four.
A more detailed description of the annotation properties can be found here.
Install the requirements with pip install -r requirements.txt
Run download_FUNSD.py to download the dataset. This will download the dataset to the datasets folder.
Run coord-to-poly.py to generate the polygon-based annotations. This will generate a new folder called datasets/FUNSD_polygon.
Run main.py to generate the augmented dataset. This will generate a new folder called datasets/FUNSD_polygon_augmented.
An optional parameter --multiplier is available to control the number of augmented images that will be generated for each input image. Defaults to 5.