README.md

metadata

pretty_name: Rule-Based Headers, FIFA World Cup 2022
task_categories:
  - video-classification
tags:
  - sports
  - football
  - soccer
  - action-spotting
  - header-detection
  - player-tracking
  - skeleton
extra_gated_heading: Request access
extra_gated_description: Internal dataset. Access is granted by hand to collaborators.
extra_gated_fields:
  Full name: text
  Institution: text
  Reason for access: text
extra_gated_button_content: Request access

Rule-Based Headers, FIFA World Cup 2022

Header annotations for all 64 matches of the 2022 FIFA Men's World Cup, together with the player and ball tracking they were derived from. Video is included for the eight knockout matches, from the quarter-finals through the final.

Seven of those eight matches have been checked by hand. Everything else is unedited detector output. If you are measuring anything, read Annotations and The final before you start.

Matches

The eight knockout matches carry video:

FIFA ID	Fixture	Round	Date	Video
128077	Croatia vs Brazil	Quarter-final	2022-12-09	M57-1.mp4
128076	Netherlands vs Argentina	Quarter-final	2022-12-09	M58-1.mp4
128079	Morocco vs Portugal	Quarter-final	2022-12-10	M59-1.mp4
128078	England vs France	Quarter-final	2022-12-10	M60-1.mp4
128080	Argentina vs Croatia	Semi-final	2022-12-13	M61-1.mp4
128081	France vs Morocco	Semi-final	2022-12-14	M62-1.mp4
128082	Croatia vs Morocco	Third place	2022-12-17	M63-1.mp4
128083	Argentina vs France	Final	2022-12-18	M64-1.mp4

Video names follow the tournament schedule, M57 being the first quarter-final and M64 the final. Directories are named by FIFA match ID.

The other 56 matches of the tournament, 128068 through 133041, are present as tracking and annotations only, without video.

Files

FIFA-headers-64games-OpenSportsLib.json
128068/                     # 56 matches, tracking only
├── live_joints.h5
└── live_ball.h5
...
128076/                     # 8 knockout matches, tracking and video
├── M58-1.mp4
├── live_joints.h5
└── live_ball.h5
...

live_joints.h5 holds frame, player_id and per-joint 3-D coordinates. live_ball.h5 holds frame, x, y, z and timestamp_utc. Missing coordinates are written as the sentinel -1.0 rather than NaN, so filter on that value before computing anything. Videos are 398x224 at 29.97 fps.

Player-centroid tracking is not included in this iteration.

Annotations

FIFA-headers-64games-OpenSportsLib.json is an OSL v2 action-spotting file with three labels: header, kickOff and finalWhistle.

Events were produced in two passes. The rule-based skeleton header spotter (h5_header_skeleton_recall_rule) proposed candidates from the joints and ball tracking, and those candidates were then reviewed and corrected by hand.

FIFA ID	Headers	Reviewed
128076	113	yes
128077	98	yes
128078	109	yes
128079	91	yes
128080	67	yes
128081	76	yes
128082	75	yes
128083	245	no

That is 629 reviewed headers across the seven checked matches.

You do not have to rely on the table above. Review status is recorded in the file: metadata.hand_verified_games lists the seven, and every match carries data[].metadata.review_status, which is either hand_verified or unreviewed.

Kick-off and final-whistle events are present for each match. A match with four kickOff events went to extra time, which is why 128076, 128077 and 128083 have four where the others have two.

The file covers all 64 matches, and every path it references now exists in this repository. The seven reviewed matches are the only reviewed ones in it; the review_status field is the reliable way to tell them apart.

The final (128083)

The 245 header events for the final are unedited detector output. They have not been checked and are not a ground truth.

A hand-annotated version of the same match exists, with 139 headers at millisecond precision:

https://github.com/drishyakarki/pixels_vs_positions/blob/main/headers_data/FWC2022-final-224p-Milli-OpenSportsLib.json

Use that file for any evaluation involving the final. The difference between 245 and 139 is roughly what the recall-tuned variant of the rule is built to produce, since it accepts false positives in order not to miss headers.

Two differences matter if you combine the two files:

• The external file records the video at 50 fps. The encode here runs at 29.97, because the annotations were made against a different encode of the same broadcast. Timings are stored in position_ms and remain correct, but converting them to frame indices needs the frame rate of the video you actually have.
• It capitalises labels (Header, Kickoff) where this file does not (header, kickOff).

Timing

Annotation positions are milliseconds into the video. Tracking rows carry UTC timestamps. The UTC_time_start field on the video input connects the two:

utc = UTC_time_start + position_ms

Do not turn a tracking row index into a time by dividing by a frame rate. Recording starts before kick-off and the clock jumps across half-time, so row index and match clock are not proportional. Read timestamp_utc instead.

Loading

from huggingface_hub import snapshot_download

# annotations only
snapshot_download("OpenSportsLab/rule-based-headers-FWC2022",
                  repo_type="dataset", allow_patterns="*.json")

# one match, all modalities
snapshot_download("OpenSportsLab/rule-based-headers-FWC2022",
                  repo_type="dataset", allow_patterns="128076/*")

# tracking for every match, no video
snapshot_download("OpenSportsLab/rule-based-headers-FWC2022",
                  repo_type="dataset", allow_patterns="*/*.h5")

Source

Video and tracking come from the PFF FC World Cup 2022 release, OpenSportsLab/PFF. The videos here are the 224p encodes, renamed from PFF game IDs to match numbers.

The header spotter that produced the proposals, its seven variants and their measured precision and recall are documented in OpenSportsLib under docs/headers/.