turbovision-miner

TurboVision Miner - Base Model

This repository contains a TurboVision miner implementation for Bittensor Subnet 44.

🎯 Model Description

This miner uses YOLOv8 models for football video analysis:

• Bounding Box Detection (football-player-detection.pt): Detects players, referees, goalkeepers, and ball
• Keypoint Detection (football-pitch-detection.pt): Detects 32 football pitch keypoints for camera calibration

📦 Repository Contents

• miner.py - Main miner implementation with model loading and inference logic
• config.yml - Chutes deployment configuration (GPU requirements, concurrency, etc.)
• football-player-detection.pt - YOLOv8x model for object detection
• football-pitch-detection.pt - YOLOv8x-pose model for keypoint detection

🚀 Deployment

This miner is designed to be deployed via the TurboVision CLI to Chutes.ai:

# Deploy to Chutes and commit on-chain
sv -vv push

# Test the deployment
sv -vv run-once

📊 Model Performance

These are base YOLO models pretrained on COCO dataset. For competitive performance:

• Fine-tune on football-specific datasets (SoccerNet, Roboflow)
• Optimize for team color classification
• Improve tracking stability across frames
• Reduce inference latency

🔧 Hardware Requirements

• GPU: A100, 3090, A40, or A6000
• VRAM: 16GB minimum
• Timeout: 300 seconds
• Concurrency: 4 instances

📝 License

This model follows the Ultralytics YOLO license.

🔗 Links

• TurboVision GitHub
• Bittensor Subnet 44
• Chutes.ai