README.md

metadata

license: other
license_name: opensportsball-bench-terms
pretty_name: OpenSportsBall benchmark data
tags:
  - video
  - sports
  - soccer
  - object-detection
  - camera-calibration
size_categories:
  - 1K<n<10K
extra_gated_heading: Request access to the OpenSportsBall benchmark data
extra_gated_description: >-
  This dataset contains video of real, identifiable adult players in soccer
  games. A maintainer reviews each request by hand, usually within a few days.
extra_gated_prompt: >-
  By requesting access you agree to these terms. Use the data only to develop or
  evaluate sports video processing, such as ball, player and pitch detection,
  camera calibration and tracking. Don't redistribute it or any frames from it,
  and don't publish frames in which a person can be recognised. Don't try to
  identify the people in it, and don't use it for face recognition, biometrics
  or surveillance. Keep your copy access-controlled and delete it when you stop
  contributing or when the maintainers ask. Access can be revoked at any time.
extra_gated_fields:
  Full name: text
  Affiliation or GitHub username: text
  What will you use the data for: text
  I agree to use this data only as described above and not to redistribute it: checkbox
extra_gated_button_content: Request access

OpenSportsBall benchmark data

The data behind the benchmarks of OpenSportsBall, an open-source system that films amateur field sports with two phones and renders a broadcast-style video. It holds the recorded video and the labels the processing pipeline is scored against, so contributors can rerun every benchmark and compare their changes against the same numbers.

Access is gated: the footage shows real adult players who can be identified, so it is shared only with contributors who accept the terms above.

Contents

The repository mirrors the footage/ directory of opensportsball-processing, so a download lands where the tools already look.

Path	What
bench/rec3/, bench/rec4cal/	The bench store for two night games: the phones' video segments the benches read (unchanged, sha256-checked against each manifest), plus sync, pitch maps, team colours, background models, play-box sets, ball label sets and detector caches
bench/fixtures/pitch/	Pitch-mapping regression fixtures and the marked-field benchmark
models/	The standard ball detector and judge, with a README of what each was trained on
rec5-game/, rec6/, rec7/	Pitch maps (pitch/) and ball label sets (labels*/, with their frames) from three more games: an orange ball, the first daylight game, and a night game with a yellow ball, with the cameras the pitch maps were made with (phase0/cameras.json). rec5-game and rec7 also have team sides (teams/), and rec7 a play-box set (playeval/). Their full videos are not included
rec8/	A fourth game outside the bench (night, 2026-11-11): its pitch map, cameras and team sides, and three play-box sets (playeval/seg0-600, seg0-3000, seg0-3600) of 10 plays the follow cam missed. From 54:37 on only Main has footage, so those panoramas show its half of the pitch. Its full video is not included

A pitch map (pitch/) holds the fitted pitch (pitch.json), the median background panorama it was fitted on (background.png, 0.03 deg per pixel) and the hand-tapped keypoints twice: taps.json in that panorama's degrees, and taps_src.json as pixels in each phone's 4K frame, which don't depend on the stitch. rec7's pitch map (since 2026-11-09) comes from a joint fit of both phones' lenses and the pitch, with no warp terms; rec5-game's and rec6's are the older stitch plus warped pitch fit.

A ball label set holds items.json (the moments and frames picked for labelling), labels.json (the ball positions, or none where it isn't visible), images/ (the labelled frames) and pano/ (the stitched frames). A set with heldout.json is kept for testing: don't train on it.

A play-box set (playeval/<set>/) holds the must-see moments of one clip that a follow cam should keep in view: clip.json (the clip), boxes.json (the reviewed box at each moment) and items.json, with the panoramas they were drawn on (pano/, pano.json).

Team sides (teams/<clip>/sides.json) say which kit components are each team and which are a keeper or referee. Where automatic side detection failed they were set by hand (sides.auto.json is what it guessed).

Getting it

After your request is approved, from a checkout of opensportsball-processing:

hf auth login                                      # a read token is enough
uv run processing/spikes/sharedata.py fetch        # everything (~37 GB)
uv run processing/spikes/sharedata.py fetch --annotations-only   # pitch maps and labels only (~1 GB)
uv run processing/spikes/benchstore.py             # check the bench video

docs/TESTING-PROCESSING.md in that repository describes each benchmark and how to run it.

Licence

The data is not under an open licence. It is shared only under the terms you accept when requesting access. The code that uses it is AGPL-3.0.