WTW Wired Table Structure Annotation

# WTW Wired Table Structure Annotation
# Based on: Rujiao Long, Wen Wang, Nan Xue, Feiyu Gao, Zhibo Yang, Yongpan
#   Wang, and Gui-Song Xia. "Parsing Table Structures in the Wild." IEEE/CVF
#   International Conference on Computer Vision (ICCV) 2021. arXiv:2109.02199.
# Paper: https://openaccess.thecvf.com/content/ICCV2021/html/Long_Parsing_Table_Structures_in_the_Wild_ICCV_2021_paper.html
# Dataset: https://github.com/wangwen-whu/WTW-Dataset
#
# Task: Annotate the cell structure of photographed and scanned wired tables
# from the WTW dataset (natural scenes, archival documents, and printed
# documents, covering seven challenging cases: inclined, curved, occluded or
# blurred, extreme aspect ratio, overlaid, multi-color, and irregular tables).
# Using the table_grid scheme, the annotator sets the table's row/column grid
# dimensions and clicks cells to mark their logical role: data, column header,
# row header, empty, or spanned (a grid position covered by a merged cell that
# starts above or to its left, which encodes the ICDAR-2019-style row/column
# start-end extent of spanning cells at the logical level). Deliberate
# simplification: WTW's full annotation also localizes each cell's physical
# quadrilateral along the inner table lines (following the ICDAR2019
# benchmark); pixel-level cell coordinates are not captured by table_grid, so
# this showcase covers the logical-structure half of the protocol only.
#
# Annotation instructions adapted from Section 3 (the WTW dataset) of the
# paper; the paper does not publish the verbatim annotator guidelines.

annotation_task_name: "WTW Wired Table Structure Annotation"
task_dir: "."

data_files:
  - sample-data.json

item_properties:
  id_key: "id"
  text_key: "image"

output_annotation_dir: "annotation_output/"
output_annotation_format: "json"

port: 8000
server_name: localhost

annotation_instructions: |
  ## Wired Table Structure Annotation

  You will see a photograph or scan of a table with visible (wired) borders,
  taken in the wild: on bottles and boxes, in archival documents, or in
  printed pages. Your job is to reconstruct the table's logical cell grid.

  ### 1. Set the grid size
  Count the rows and columns of the table as defined by its inner table lines
  (following the ICDAR2019 convention, cell extents are defined by the inner
  lines, not the outer border). Set the grid dimensions accordingly. A cell
  that spans several rows or columns still occupies all of those grid
  positions.

  ### 2. Mark cell roles
  Click each cell to cycle through its role:

  - **data** -- an ordinary content cell (the default; you do not need to
    mark these explicitly).
  - **col_header** -- a cell that labels the column(s) below it.
  - **row_header** -- a cell that labels the row(s) to its right.
  - **empty** -- a cell position with no content at all.
  - **spanned** -- a grid position covered by a merged cell that starts above
    it or to its left. For a cell spanning columns 2-3, mark position 2 with
    the cell's real role and position 3 as spanned; likewise for row spans.
    This records each spanning cell's row/column start and end extent.

  ### 3. Challenging cases
  Select every challenging condition that applies to this table image:
  inclined, curved, occluded or blurred, extreme aspect ratio, overlaid,
  multi-color, or irregular. Select none if the table is a plain upright
  scan.

  ### 4. Image source
  Finally, record whether the image is a natural-scene photograph, an
  archival document, or a printed document.

  Annotate the table exactly as printed, even if it contains errors. If the
  image contains more than one table, annotate the largest fully visible one
  and skip the item if that is ambiguous.

annotation_schemes:
  - annotation_type: table_grid
    name: structure
    description: "Set the grid size, then click cells to mark headers, empty cells, and spanned continuations."
    image_key: image
    rows_key: rows
    cols_key: cols
    roles: [data, col_header, row_header, empty, spanned]

  - annotation_type: multiselect
    name: challenging_cases
    description: "Select all challenging conditions that apply to this table image (per the WTW taxonomy)."
    labels:
      - name: "inclined"
        tooltip: "The table is rotated or tilted relative to the image frame"
        key_value: "1"
      - name: "curved"
        tooltip: "The table surface is bent or warped, e.g., on a bottle or curved page"
        key_value: "2"
      - name: "occluded-or-blurred"
        tooltip: "Parts of the table are hidden by objects, glare, or blur"
        key_value: "3"
      - name: "extreme-aspect-ratio"
        tooltip: "The table is unusually wide or tall relative to its other dimension"
        key_value: "4"
      - name: "overlaid"
        tooltip: "Text, stamps, or graphics are printed over the table lines"
        key_value: "5"
      - name: "multi-color"
        tooltip: "Cells or lines use multiple colors that complicate line detection"
        key_value: "6"
      - name: "irregular"
        tooltip: "Non-rectangular layout with unusual cell arrangements"
        key_value: "7"

  - annotation_type: radio
    name: image_source
    description: "What kind of image is this table from?"
    labels:
      - name: "natural-scene"
        tooltip: "A photograph of a table in the wild, e.g., on packaging, signs, or labels"
        key_value: "q"
      - name: "archival-document"
        tooltip: "A scan or photograph of a historical/archival document"
        key_value: "w"
      - name: "printed-document"
        tooltip: "A scan or screenshot of a modern printed or digital document"
        key_value: "e"

allow_all_users: true
instances_per_annotator: 25
annotation_per_instance: 2
allow_skip: true