README.md

metadata

license: apache-2.0
task_categories:
  - object-detection
tags:
  - document-intelligence
  - form-field-detection
  - pdf
  - multilingual
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
      - split: val
        path: data/val-*
dataset_info:
  features:
    - name: id
      dtype: int64
    - name: image_id
      dtype: int64
    - name: image
      dtype: image
    - name: file_name
      dtype: string
    - name: width
      dtype: int64
    - name: height
      dtype: int64
    - name: objects
      sequence:
        - name: id
          dtype: int64
        - name: area
          dtype: float32
        - name: bbox
          sequence: float32
        - name: category
          dtype:
            class_label:
              names:
                '0': Text
                '1': CheckBox
                '2': Signature
        - name: category_id
          dtype: int64
        - name: iscrowd
          dtype: int64
        - name: segmentation
          sequence: int64
  splits:
    - name: train
      num_bytes: 384892618937.625
      num_examples: 435713
    - name: test
      num_bytes: 25280639959.69
      num_examples: 33061
    - name: val
      num_bytes: 9380337961.835
      num_examples: 18195
  download_size: 308455257247
  dataset_size: 419553596859.15

CommonForms: A Large, Diverse Dataset for Form Field Detection

This repository hosts the CommonForms dataset, a web-scale dataset for form field detection, introduced in the paper CommonForms: A Large, Diverse Dataset for Form Field Detection.

CommonForms casts the problem of form field detection as object detection: given an image of a page, predict the location and type (Text Input, Choice Button, Signature) of form fields.

Key Features:

• Scale: Roughly 55,000 documents comprising over 450,000 pages.
• Source: Constructed by filtering Common Crawl to find PDFs with fillable elements.
• Diversity: Contains a diverse mixture of languages (one third non-English) and domains, with no single domain making up more than 25% of the dataset.
• Purpose: The first large-scale dataset released for form field detection, aimed at fostering the development of robust form field detectors.

Paper: CommonForms: A Large, Diverse Dataset for Form Field Detection Code: https://github.com/jbarrow/commonforms