💷📊
Computer Vision PipelineExperimental

Video Analysis

Upload a broadcast clip → detect players & ball → track across frames → project to pitch coordinates → spot events. Download JSON/CSV.

Pipeline Architecture
End-to-end broadcast video → structured tracking + event data
Upload
Detection
Tracking
Calibration
Projection
Events

Drop a broadcast clip here

or click to browse. Supports MP4, WebM, MOV, AVI (max 500MB)

or
Known Limitations
Camera Cuts

Broadcast footage has frequent camera changes. Calibration and tracking reset per shot — expect gaps in tracking during cuts.

Player Identity

Shirt number / re-ID is a separate task. Track IDs are consistent within a shot but not across camera changes.

Ball Tracking

Ball visibility is intermittent in broadcast. The ball tracker is the most fragile component — expect missing detections.

Processing Time

Full pipeline processing can take 2-10× the clip duration depending on resolution. Short clips (<30s) recommended.

Pipeline Components
Open-source models and tools used in each stage
Detection

Soccer-tuned YOLO model from Hugging Face. Detects players, goalkeepers, referees, and ball with bounding boxes.

Tracking

ByteTrack for multi-object tracking. Assigns consistent IDs across frames within each camera shot.

Calibration

SoccerNet camera calibration tooling. Estimates homography to map pixel space → pitch coordinates.

Action Spotting

SoccerNet pretrained action spotting models. Detects goals, cards, substitutions, and 17 action classes.

Ball Actions

SoccerNet BAS baselines for dense on-ball events: passes, dribbles, duels — closer to an event stream.

Export

Tracking: frame, timestamp, track_id, class, confidence, bbox, pitch_x, pitch_y. Events: timestamp, label, confidence.