README.md

metadata

dataset_info:
  features:
    - name: image
      dtype: image
    - name: filename
      dtype: string
    - name: subtype_label
      dtype: string
    - name: modifier
      dtype: bool
    - name: transcription
      dtype: string
    - name: metadata_json
      dtype: string
    - name: series_title
      dtype: string
    - name: episode_no
      dtype: string
    - name: program_title
      dtype: string
    - name: episode_title
      dtype: string
    - name: director
      dtype: string
    - name: producer
      dtype: string
  splits:
    - name: train
      num_bytes: 8277937
      num_examples: 452
    - name: test
      num_bytes: 968768
      num_examples: 51
  download_size: 8981265
  dataset_size: 9246705
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
task_categories:
  - image-to-text
language:
  - en
size_categories:
  - n<1K

Transcribed Slates

General Information

This dataset was created for the training and testing of machine learning systems for extracting information from slates/on-screen or filmed text in video productions. The data associated with each instance was acquired by observing text on the slates in the file. There are two levels of data collected, a direct transcription and contextual information. For the direct transcription if there was illegible text an approximation was derived. The information is reported by the original creator of the slates and can be assumed to be accurate.

The data was collected using a software made specifically to categorize and transcribe metadata from these instances (see file directory description). The transcription was written in a natural reading order (for a western audience), so right to left and top to bottom. If the instance was labeled “Graphical” then the reading order was also right to left and top to bottom within individual sections as well as work as a whole.

This dataset was created by Madison Courtney, in collaboration with GBH Archives staff, and in consultation with researchers in the Brandeis University Department of Computer Science.

Uniqueness and overlapping data

Some of the slates come from different episodes of the same series; therefore, some slates have data overlap. For example, the “series-title” may be common across many slates. However, each slate instance in this dataset was labeled independently of the others. No information was removed, but not every slate contains the same information.

Different “sub-types” of slates have different graphical features, and present unique challenges for interpretation. In general, sub-types H (Handwritten), G (Graphical), C (Clapperboard) are more complex than D (Simple digital text) and B (Slate over bars). Most instances in the dataset are D. Users may wish to restrict the set to only those with subtype D.

Labels and annotations were created by an expert human judge. In Version 2, labels and annotations were created only once without any measure of inter-annotator agreement. In Version 3, all data were confirmed and/or edited by a second expert human judge. The dataset is self-contained. But more information about the assets from which these slates were taken can be found at the main website of the AAPB https://www.americanarchive.org/

Data size and structure

The data is tabular. There are 7 columns and 503 rows. Each row represents a different labeled image. The image files themselves are included in the dataset directory. The columns are as follows:

• 0: filename : The name of the image file for this slate
• 1: seen : A boolean book-keeping field used during the annotation process
• 2: type-label : The type of scene pictured in the image. All images in this set have type "S" signifying "Slate"
• 3: subtype-label : The sub-type of scene pictured in the image. Possible subtypes are "H" (Handwritten), "C" (Clapperboard), "D" (Simple digital text), "B" (Slate over bars), "G" (Graphical).
• 4: modifier : A boolean value indicating whether the slate was "transitional" in the sense that the still image was captured as the slate was fading in or out of view.
• 5: note-3 : Verbatim transcription of the text appearing on the slate
• 6: note-4 : Data in key-value structure indicating important data values presented on the slate. Possible keys are "program-title", "episode-title", "series-title", "title", "episode-no", "create-date", "air-date", "date", "director", "producer", "camera". Dates were normalized as YYYY-MM-DD. Names were normalized as Last, First Middle.

Data format

The directory contains the tabular data, the image files, and a small utility for viewing and/or editing labels. The Keystroke Labeler utility is a simple, serverless HTML-based viewer/editor. You can use the Keystroke Labeler by simply opening labeler.html in your web browser. The data are also provided serialized as JSON and CSV. The exact same label data appears redundantly in these 3 files:

• img_arr_prog.js - the label data loaded by the Keystroke Labeler
• img_labels.csv - the label data serialized as CSV
• img_labels.json - the label data serialized as JSON

This dataset includes metadata about programs in the American Archive of Public Broadcasting. Any use of programs referenced by this dataset are subject to the terms of use set by the American Archive of Public Broadcasting.