wunderscout 0.1.19

Scouting and vision tools for YOLO and sports analytics.

wunderscout

A Python library for extracting player and ball tracking data from soccer match footage using YOLO, Siglip embeddings, and homography.

Features

• Detection & Tracking: Uses YOLO for player/ball/pitch-keypoint detection and ByteTrack for temporal consistency.
• Automated Team Clustering: Groups players into teams using Siglip vision transformer embeddings and K-Means clustering via UMAP dimensionality reduction.
• Pitch Mapping: Transforms 2D image coordinates to a normalized 0-1 coordinate system using pitch keypoint homography.
• Goalkeeper Attribution: Assigns goalkeepers to teams based on proximity to team centroids.
• Data Export: Generates Team1 and Team2 CSV files containing frame-by-frame XY coordinates.
• Heatmap Generation: Creates histogram and KDE-based spatial density maps for individual players and teams.

Installation

pip install wunderscout

Quick Start

from wunderscout import Detector, DataExporter, HeatmapGenerator

# Initialize with paths to trained YOLO weights
detector = Detector(
    player_weights="players.pt",
    field_weights="pitch.pt"
)

# Run processing
result = detector.run(
    video_path="input_match.mp4",
    output_video_path="output_match.mp4"
)

# Export tracking data to CSV
exporter = DataExporter()
exporter.save_csvs(result, "output/tracking.csv")

# Generate heatmaps
heatmap_gen = HeatmapGenerator()
player_heatmap = heatmap_gen.generate_player_heatmap(result, player_id=5, method="both")
heatmap_gen.save_heatmap(player_heatmap, "output/player_5.json")

Core Components

Detector

Main pipeline coordinator that orchestrates detection, tracking, and classification.

Parameters:

• player_weights (str): Path to YOLO player detection model
• field_weights (str): Path to YOLO field keypoint detection model

Methods:

• run(video_path, output_video_path=None): Process video and return TrackingResult

Models

Handles YOLO detection and SiGLIP embedding extraction.

Key Methods:

• detect_players(frame, conf=0.3): Detect players, goalkeepers, referees, and ball
• detect_field(frame, conf=0.3): Detect pitch keypoints for homography
• get_embeddings(pil_crops, batch_size=32): Extract visual embeddings for team classification
• get_calibration_crops(video_path, stride=30): Sample player crops for calibration

Detection Class IDs:

• 0: Ball
• 1: Goalkeeper
• 2: Player
• 3: Referee

TeamClassifier

Clusters players into teams using UMAP + K-Means with temporal smoothing.

Methods:

• fit(embeddings): Train clustering model on calibration embeddings
• get_consensus_team(tracker_id, embedding): Assign team ID with 50-frame rolling consensus
• resolve_goalkeepers_team_id(players, goalkeepers): Assign goalkeepers by centroid proximity
• get_final_assignments(): Get final team assignments (dict: tracker_id → team_id)

PitchMapper

Projects pixel coordinates to normalized pitch coordinates [0, 1] × [0, 1] using homography.

Methods:

• get_matrix(keypoints_xy, keypoints_conf): Compute homography from detected pitch keypoints (requires ≥4 keypoints with conf > 0.5)
• transform(points, H=None): Transform points to normalized pitch coordinates

HeatmapGenerator

Generates spatial density maps using histogram binning and/or Gaussian KDE.

Parameters:

• pitch_length (float): Pitch length in meters (default: 105.0)
• pitch_width (float): Pitch width in meters (default: 68.0)
• histogram_bins (tuple): Histogram resolution (default: (50, 34))
• kde_grid_size (tuple): KDE grid resolution (default: (100, 68))
• min_samples_for_kde (int): Minimum samples for KDE (default: 10)

Methods:

• generate_player_heatmap(result, player_id, method="both"): Generate heatmap for one player
• generate_team_heatmap(result, team, method="both"): Generate aggregated team heatmap
• generate_all_players_heatmaps(result, method="both"): Generate heatmaps for all players
• save_heatmap(heatmap_data, output_path, pretty=False): Save heatmap to JSON

DataExporter

Exports tracking data to CSV format.

Methods:

• save_csvs(result, output_path): Export tracking data (creates {base_name}_Team1.csv and {base_name}_Team2.csv)

TrackingResult

Data class containing tracking results.

Attributes:

• frames (dict): Frame-by-frame tracking data {frame_idx: {"players": {player_id: (x, y)}, "ball": (x, y) or None}}
• team_assignments (dict): Player ID → Team ID mapping
• total_frames (int): Total number of processed frames
• fps (float): Video frame rate

Methods:

• get_team_players(team): Get player IDs for team 0 or 1
• get_all_player_ids(): Get all tracked player IDs
• get_player_trajectory(player_id): Get position history for one player
• get_ball_trajectory(): Get ball position history

Training Custom Models

from wunderscout import Trainer

trainer = Trainer(api_key="your_roboflow_api_key")

# Train player detection model
trainer.train_players(
    workspace="soccer-analytics",
    project="player-detection",
    version=1,
    epochs=300,
    output_dir="runs/training/player"
)

# Train field keypoint model
trainer.train_field(
    workspace="soccer-analytics",
    project="field-keypoints",
    version=1,
    epochs=300,
    output_dir="runs/training/field"
)

Output Formats

CSV Export

Two CSV files (one per team) with structure:

,,,Team1,Team1,Team1,Team1,,
,,,3,3,7,7,,
Period,Frame,Time [s],Player3_X,Player3_Y,Player7_X,Player7_Y,Ball_X,Ball_Y
1,0,0.00,0.234,0.567,0.789,0.345,0.500,0.500
1,1,0.04,0.235,0.568,NaN,NaN,0.501,0.499

Coordinate System:

• Range: [0, 1] × [0, 1]
• (0, 0) = Top-left corner of pitch
• (1, 1) = Bottom-right corner of pitch
• Missing data: "NaN"

Heatmap JSON

{
  "player_id": 5,
  "sample_count": 1500,
  "histogram": {
    "xedges": [0.0, 2.1, 4.2, ..., 105.0],
    "yedges": [0.0, 2.0, 4.0, ..., 68.0],
    "values": [[count_00, count_01, ...], [count_10, count_11, ...], ...]
  },
  "kde": {
    "x": [0.0, 1.05, 2.1, ..., 105.0],
    "y": [0.0, 1.0, 2.0, ..., 68.0],
    "values": [[density_00, density_01, ...], [density_10, density_11, ...], ...]
  }
}

Pitch Coordinate System

Normalized coordinate system mapping to standard football pitch (105m × 68m).

Coordinate Range:

• X-axis: 0.0 (left goal line) → 1.0 (right goal line)
• Y-axis: 0.0 (top sideline) → 1.0 (bottom sideline)

Key Landmarks:

Landmark	X	Y	Description
Left goal (top post)	0.000	0.365	Top edge of left goal
Left goal (bottom post)	0.000	0.635	Bottom edge of left goal
Left penalty spot	0.105	0.500	11m from goal line
Center circle	0.500	0.500	Pitch center
Right penalty spot	0.895	0.500	11m from goal line
Right goal (top post)	1.000	0.365	Top edge of right goal
Right goal (bottom post)	1.000	0.635	Bottom edge of right goal

Dependencies

• ultralytics
• supervision
• transformers
• umap-learn
• scikit-learn
• opencv-python
• scipy
• more-itertools
• torch
• numpy
• roboflow

License

MIT