⚡ Ball Action Spotting (BAS)

Task Overview

Input: Panoramic 4K video (full match)
Output: Temporal action labels for 12 ball event classes with global timestamps

Input / Output Definition

Input

• videos/match_XXXXX.mp4 — Full match panoramic video (4K)

Output (per action)

12 Action Classes

File Layout

SoccerTrack-v2/
├── videos/
│   ├── 117093/
│   │   ├── 117093_panorama_1st_half.mp4
│   │   └── 117093_panorama_2nd_half.mp4
│   └── ...
├── gsr/
├── bas/
│   ├── 117093/
│   │   └── 117093_12_class_events.json
│   └── ...
├── mot/
└── raw/
    └── ...

Each match ships one BAS annotation file covering both halves. The gameTime field encodes which half each event belongs to.

See format-bas.md for the full schema reference.

Schema & SoccerNet Mapping

Example JSON Structure

{
  "UrlLocal": "117093",
  "UrlYoutube": null,
  "annotations": [
    {
      "gameTime": "1 - 12:30",
      "position": "750000",
      "label": "Pass",
      "team": "left",
      "player_id": "117093_L_9",
      "visibility": "visible"
    },
    ...
  ]
}

position is milliseconds from kickoff of the half encoded in gameTime, not from match start. Convert to a frame index with round(int(position) / 40) at 25 fps.

SoccerNet BAS Correspondence

Evaluation / Metrics

Evaluation uses tolerant mean Average Precision, following the SoccerNet Ball Action Spotting Challenge. Report two tolerances:

• mAP@1s — tight (precise temporal localisation)
• mAP@5s — loose (event-level detection)

Reference CLI:

python -m src.evaluation.bas_map \
    --pred preds/bas --gt data/bas --matches 117099 117100 \
    --tolerances 1 5

Starter Kit

Differences vs SoccerNet BAS