Offside

Work in progress. Core pipeline is functional but the UI needs significant work. Not production-ready.

Computer vision and tactical analytics for broadcast soccer footage.

Live demo: coming soon

What it does

Upload or select a broadcast clip. The pipeline:

1. Detects players, goalkeepers, referees, and the ball using YOLOv8m
2. Tracks each entity across frames with ByteTrack (persistent IDs)
3. Maps pixel positions to real pitch coordinates via homography estimation
4. Assigns teams via KMeans clustering on torso HSV colors
5. Computes per-frame formation (3- or 4-line KMeans on pitch depth)
6. Computes pressing intensity (0–100 score based on nearest opponents to ball carrier)
7. Computes Voronoi space control (% of pitch controlled per team)
8. Aggregates everything into a tactical dashboard

Stack

Layer	Tech
Detection	YOLOv8m (Ultralytics)
Tracking	ByteTrack via boxmot
Homography	OpenCV (HSV line detection + RANSAC)
Analytics	NumPy, SciPy, scikit-learn, Shapely
Backend	FastAPI + SSE streaming
Frontend	React 18 + Vite, plain CSS modules
Deployment	HuggingFace Spaces (backend) + Vercel (frontend)

Running locally

Backend

cd backend
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
uvicorn app.main:app --reload --port 8000

Frontend

cd frontend
npm install
npm run dev   # http://localhost:5173

Fine-tune the detector (optional)

cd backend
python model/train.py --epochs 50 --batch 16

Demo clips

Five broadcast clips are pre-loaded in backend/demo_clips/:

ID	Match
possession	Possession sequence
pressing	High-press sequence
open_play	Chelsea vs Arsenal
transition	Tottenham vs Watford
ucl	Liverpool vs PSG (UCL)

Deployment

• Backend: HuggingFace Spaces (Docker) — see backend/Dockerfile
• Frontend: Vercel — set VITE_API_URL to your HF Space URL

Work in progress