SkillCorner/opendata-bodypose

SkillCorner Open Data — Body Pose

3D body-pose data derived from broadcast video, released alongside the SkillCorner Open Data repository as a joint initiative between SkillCorner and PySport.

Initial testing release. Two matches, published so the community can work with the format and tell us what is useful before we consider a wider release. Feedback is genuinely wanted — open an issue on the opendata repo or reply in the Community tab here.

What is in here

File	Match	Size
raw/1925299.jsonl.zip	Brisbane Roar v Perth Glory, 2024-12-21	595 MB
raw/1996435.jsonl.zip	Sydney FC v Adelaide United, 2025-02-01	614 MB

Each archive contains one .jsonl file — one JSON object per video frame, about 3.3 GB uncompressed. Both matches are from the Australian A-League 2024/2025 season (competition edition 870), and both already have tracking, dynamic events and phases of play published in the opendata repo, so pose can be joined to the rest of the data for the same games.

Record structure

One object per frame:

{
  "frame": 2438,
  "timestamp": "00:01:36.52",
  "period": 1,
  "ball_data": {"x": 0.35, "y": -22.22, "z": 7.074, "is_detected": true},
  "possession": {"player_id": null, "group": "away team"},
  "image_corners_projection": {"x_top_left": -44.26, "y_top_left": 39.0, "...": "..."},
  "player_data": [
    {
      "player_id": 809166,
      "x": -17.36,
      "y": -22.06,
      "is_detected": true,
      "joints": {
        "lAnkle": {"xyz": [-25.042, 0.86, 0.204], "p90_mae_cm": 12.39},
        "lEar":   {"xyz": [-25.016, 0.815, 1.809], "p90_mae_cm": 9.29},
        "...": "..."
      }
    }
  ]
}

Coordinates are in metres, origin at the centre of the pitch, x along the long axis and y along the short axis — the same convention as the tracking data in the opendata repo.

z is not a height above the pitch. It is accurate relative to the player's centroid, but not necessarily in the pitch coordinate system, so it can be negative. Treat it as pose geometry rather than elevation. This will improve with future Body Pose enhancements.

Note that ball_data also carries z, which the 10 fps tracking files do not.

The 2D x/y positions in this file are the standard tracking positions interpolated up to 25 fps, not native 25 fps measurements.

The 29 joints

nose, neck, lEye, rEye, lEar, rEar, lShoulder, rShoulder, lElbow, rElbow, lWrist, rWrist, lThumb, rThumb, lPinky, rPinky, midHip, lHip, rHip, lKnee, rKnee, lAnkle, rAnkle, lHeel, rHeel, lBigToe, rBigToe, lSmallToe, rSmallToe

p90_mae_cm

Every joint carries its own accuracy estimate: the 90th-percentile mean absolute error in centimetres. Typical values run 7–15 cm, varying by joint — torso and head landmarks are more reliable than extremities. Use it to filter: joints with a high p90_mae_cm should not be trusted for fine-grained biomechanics.

This also means positions are reported to millimetre precision but are only accurate to a handful of centimetres. Do not read more into the decimals than the error estimate supports.

Frame rate — read this before joining to tracking

Pose is 25 fps. The tracking data in the opendata repo is 10 fps. The frame numbering differs accordingly: this file runs to frame 153,252 for a match whose tracking file ends at frame 61,300.

Both cover the same wall-clock time, so:

pose_frame = 2.5 * tracking_frame

Every 5th pose frame (a multiple of 5) lands exactly on an even-numbered tracking frame. Those are the frames where pose and tracking align without interpolation.

Coverage

Pose is not available for every player in every frame — it requires the player to be visible and resolvable in the broadcast image. For match 1925299:

Frames in file	153,252
Frames with at least one pose	106,757 (69.7%)
Player-frames total	4,902,560
Player-frames with pose	1,540,715 (31.4%)
Average players with pose, per frame that has any	14.4
Total joint observations	44,680,735

A player's joints field is null when pose could not be resolved for them in that frame, while their x/y position may still be present. Expect gaps when players are off-camera, occluded, or in a crowd.

Limitations

• Body Pose is only provided for detected players.
• Position on the z-axis is accurate relatively to the player's centroid, but not necessarily in the pitch coordinate system. This will come with future Body Pose enhancements.
• XY Tracking data and Body Pose are generated separately. This can cause minor misalignments between the XY attributes (positions, is_detected flag) and the Body Pose joints.

Loading

The archives are large, so download rather than stream:

import json
import zipfile

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="SkillCorner/opendata-bodypose",
    filename="raw/1925299.jsonl.zip",
    repo_type="dataset",
)

with zipfile.ZipFile(path) as z:
    with z.open("1925299.jsonl") as fh:
        for line in fh:
            frame = json.loads(line)
            for player in frame["player_data"]:
                if player["joints"]:
                    ...  # 29 joints, each with xyz and p90_mae_cm

Reading the whole match into memory as JSON is not advisable — process it frame by frame, or convert the parts you need into a columnar format first.

Related data

The rest of the open dataset — broadcast tracking, dynamic events, phases of play and season aggregates for 20 A-League matches — lives at github.com/SkillCorner/opendata, along with tutorials.

Licence and credit

MIT, matching the opendata repository.

If you use this data we kindly ask that you credit SkillCorner, and we would love to hear what you build — tell us on X/Twitter.