royshilkrot/scoreboard-fields

Scoreboard Fields

291 photos of real scoreboards, with a box around each field on the board (home score, away score, clock, period, and the other fields) and the four corners of each board.

The labels were made for ScoreSight, an app that reads scoreboards from a video feed with OCR. Setting a board up means drawing a box on each field and marking the board's corners so the picture can be straightened. This dataset is for training models that suggest both.

Contents

split	photos	home	away	clock	period	other	board quads
train	189	140	133	107	31	1,178	188
validation	30	22	23	20	5	181	27
test	72	60	55	42	11	330	72

• 232 photos from Wikimedia Commons and 59 from Open Images V7 (its "Scoreboard" class, Flickr photos). Association football, cricket, tennis, American football and baseball are the most common sports, and there are a few photos of athletics, table tennis, Australian rules football, curling and wrestling.
• Boards of every kind: lit 7-segment and dot-matrix digits, LED video boards, flip cards on cricket pavilions, TV graphics.
• 11 photos have no board quad (the board's corners are out of the frame or unclear), and 7 have more than one board. 56 photos have only other fields: boards with no home/away score, clock or period, such as a cricket total or a results table.

Format

Each split is a folder of JPEG images and a metadata.jsonl (the Hub's image folder format), one row per photo:

{
  "file_name": "cm10219080.jpg",
  "id": "cm10219080",
  "width": 1280, "height": 917,
  "objects": {"bbox": [[336.3, 410.2, 41.2, 46.1], ...], "category": ["home", ...]},
  "boards": [[[157.6, 382.9], [763.0, 468.8], [762.6, 553.0], [144.8, 486.9]]],
  "sport": "association football",
  "source": "wikimedia-commons",
  "source_url": "https://commons.wikimedia.org/wiki/File:...",
  "author": "...",
  "license": "CC-BY-SA-4.0"
}

• bbox is [x, y, width, height] in pixels of the image as provided here (COCO style).
• category is one of:
  • home: the home team's score (or the first team's, when the board doesn't say which is home)
  • away: the other team's score
  • clock: the game clock
  • period: the period, quarter, half, inning or set number
  • other: any other field (team names, fouls, timeouts, sets and games, runs and wickets, players' numbers, labels)
• boards lists each board's corners as [x, y] points in the order top left, top right, bottom right, bottom left.
• sport comes from the photo's Commons categories. It is empty for the Open Images photos and for some Commons ones.

The same labels are in COCO detection format in annotations/{train,validation,test}.json (category ids 0–4 in the order above; each image entry also has its boards).

from datasets import load_dataset

ds = load_dataset("royshilkrot/scoreboard-fields")
row = ds["train"][0]
row["image"], row["objects"]["bbox"], row["objects"]["category"], row["boards"]

How it was made

1. Candidates: Wikimedia Commons photos found by sport, and Open Images photos of the Scoreboard class whose board covers at least 2% of the photo. Each was looked at on a contact sheet, and the usable ones kept.
2. Suggestions: each photo was pre-labelled automatically. Boxes came from a PP-OCRv6 text detector plus lit-digit proposals read by PP-OCR, and roles from a small model.
3. Labelling: every photo was then reviewed by one person (the dataset's author) on a web page. They corrected, added and removed boxes, set the roles, and drew the board corners by hand. A second pass removed the wrong other boxes.

The split is by a hash of the photo id: about 2/8 test, 1/8 validation, the rest train. Photos already used to train an earlier model are in train. ScoreSight's own OCR evaluation set (186 other Commons photos) is not included, so it stays unseen.

Limitations

• One labeller. The other boxes cover the fields that matter on each board, but not every piece of text on it.
• period is rare (47 boxes).
• Many photos are wide shots where the board is small. Broadcast graphics and close-ups are a minority.
• The images were resized and recompressed for labelling (long side 640–1950 px).

Licence

• Labels (boxes, roles, board corners, the files in annotations/ and the label fields of metadata.jsonl): CC BY 4.0, © Roy Shilkrot.

• Images: each photo keeps the licence its author gave it. That licence and the author are given with every row and in attribution.csv, with a link to the original:

  licence	photos
  CC BY-SA 4.0	106
  CC BY 2.0	88
  CC BY-SA 2.0	29
  CC BY-SA 3.0	26
  CC0 1.0	16
  CC BY 3.0	10
  Public Domain Mark 1.0	8
  CC BY 4.0	7
  CC BY 2.5	1

  The only changes to the photos are resizing and JPEG recompression. If you redistribute them, credit each author as attribution.csv lists them, and keep the share-alike photos under their licence.