GSR Task - SoccerTrack v2

Task Overview

Input: Panoramic 4K video (full match)
Output: Per-frame game state on 2D pitch minimap (position + role + team + jersey number)

                    Alignment with SoccerNet-GSR:

                    ✅ Same: 2D pitch coordinates, player roles, team assignment, jersey numbers

                    ⚠️ Different: Fixed panoramic camera + full match (not broadcast clips)

Input / Output Definition

Input

videos/match_XXXXX.mp4 — Full match panoramic video (4K)
Camera: BePro Cerberus (2 matches) or 3-camera panoramic (8 matches)

Output (per frame)

Field	Type	Description
`frame`	int	Frame number
`time`	float	Timestamp in video (seconds)
`player_id`	int	Unique track ID (persistent throughout match)
`x, y`	float	2D pitch coordinates in meters
`role`	string	player / goalkeeper / referee / other
`team_side`	string	left / right / null
`jersey_number`	int/null	0–99 or null

File Layout

SoccerTrack-v2/
├── videos/
│   ├── 117093/
│   │   ├── 117093_panorama_1st_half.mp4
│   │   └── 117093_panorama_2nd_half.mp4
│   └── ...
├── gsr/
│   ├── 117093/
│   │   ├── 117093_1st.json
│   │   └── 117093_2nd.json
│   └── ...
├── bas/
├── mot/
└── raw/
    ├── 117093/
    │   ├── 117093_keypoints.json
    │   ├── 117093_mapx.npy
    │   ├── 117093_mapy.npy
    │   └── ...
    └── ...

Each match ships one GSR annotation file per half. Unlike SoccerNet-GSR which stores image frames under img1/, we reference the trimmed half videos directly from videos/ to save storage space.

See format-gsr.md for the full schema reference and pitch coordinate system.

Schema & SoccerNet Mapping

Example JSON Record

Each half's file is a flat JSON array of per-entity-per-frame records. One record:

{
  "image_id": 12480,
  "track_id": 7,
  "player_id": "117092_L_9",
  "role": "player",
  "jersey_number": 9,
  "team_side": "left",
  "x": 48.21,
  "y": 34.07,
  "bbox_image": [1840, 710, 108, 242],
  "bbox_pitch":  [46.8, 33.1, 1.4, 2.0]
}

SoccerNet-GSR Correspondence

SoccerTrack v2	SoccerNet-GSR	Notes
`x, y`	`x_pos, y_pos`	Pitch coordinates (meters)
`team_side`	`left / right`	Team assignment
`jersey_number`	`jersey_number`	0–99
`role`	`role`	player/goalkeeper/referee/other
`player_id`	`track_id`	Persistent across full match

Evaluation / Metrics

Evaluation uses GS-HOTA (Game State Higher Order Tracking Accuracy), following the SoccerNet-GSR Challenge definition. Report GS-HOTA plus the DetA / AssA / LocA decomposition.

Reference CLI:

python -m src.evaluation.gs_hota \
    --pred preds/gsr --gt data/gsr --matches 117099 117100

The module is a thin wrapper over the upstream sn-gamestate scorer.

Starter Kit

Get running fast. The GSR starter kit scaffolds detection → tracking → jersey/role tagging → pitch projection into a runnable pipeline on one GPU. Fill in the model calls, then run eval.

📦 baselines/gsr/ 🏁 Leaderboard

Differences vs SoccerNet-GSR

Key Differences:

Camera view: Panoramic (full-pitch) vs. broadcast (partial view with occlusions)
Duration: Full match (~90 min) vs. short clips
Camera setup: Fixed BePro/3-camera system vs. moving broadcast camera
Level: University amateur matches vs. professional leagues