doc-ufcn utilities.

In this repository, we provide the tools to manipulate the predictions made by a segmentation model that outputs segmentation masks. The segmentation are all merged into one file as described in the image below. Here, each colors matches a different class: green is for text line, blue for page header, orange for page footer and red for table headers. This is the segmentation mask of a census table image.

The provides provides two utilities:

• convert-mask-to-docufcn-polygon: this is used to convert a simple image segmentation mask to a .json file that describes the segmentation, as detailed in Data and configuration.
• resize-docufcn-polygons: with this tool, you can reshape the segmentation of a .json file. It is used in the context where you trained and tested the neural network on small images but would like to upscale the resolution in order to match the original image size.

Installation

Data and configuration

The JSON files that contain the segmentations have the following structure:

{
  "img_size": [
    height,
    width
  ],
  "class1": [
    {
      "confidence": confidence,
      "polygon": [...]
    }
  ],
  "class2": [
    {
      "confidence": confidence,
      "polygon": [...]
    },
    {
      "confidence": confidence,
      "polygon": [...]
    }
  ]
}

And the JSON file used to identify each segmentation class from the image color has the following structure:

[
  {
    "id": 1,
    "name": "text_line",
    "color": [
      0,
      153,
      0
    ]
  }
]

Command line interface