Adit-jain/soccana

⚽ SoccerNet Object Detection Model (YOLOv8)

This repository hosts a YOLOv8-based object detection model trained on a curated and segmented dataset derived from SoccerNet and other public football datasets. The model is designed to detect key entities in a football game — players, referees, and the ball — with high accuracy, even in challenging scenes.

---

🚀 Overview

• Architecture: YOLOv8n + SAHI (Sliced Aided Hyper Inference)
• Input Size: 1280x1280
• Epochs Trained: 200
• Batch Size: 32
• Classes:
  • Player
  • Referee
  • Ball
• Formats:
  • [✓] YOLOv8 format (Ultralytics)
  • [✓] COCO format (JSON)

---

Demo

SAMPLE LINK : DRIVE

Note : This sample uses Kmeans, UMAP and SIGLIP for team assignment. This does not have Re-identification applied, hence the large player numbers.

---

🧠 Capabilities

• Robust detection of small objects like the ball across varied resolutions (160x160 to 1280x1080).
• Works well on both standard game footage and camera-angled shots.
• Can be paired with:
  • ⚽ KMeans clustering for team classification (with jersey color embeddings from SIGLIP).
  • 📈 Ball interpolation for smooth ball trajectory tracking.
  • 🧠 Future integration with ByteTrack or BoT-SORT for player tracking.

---

📦 Training Details

Parameter	Value
Epochs	200
Batch Size	32
Image Size	1280
Optimizer	Auto
Pretrained	True
Seed	44
Det. Inference	True
Dropout	0.3
Patience	100 (early stop)
iOU Threshold	0.7
Augmentations	RandAugment + Erasing
AutoAugment	Enabled
Mosaic	On
Flip LR	50%

---

🧾 How to Use

A detailed guide and code can be found at github

🧊 Notes

Model supports sliced image inference using SAHI, optimized for high-resolution input. Ideal for sports analytics, heatmaps, player positioning, and advanced tracking systems.