Custom YOLOv8 model built from real gameplay

🧠 Project Overview

Over the course of one day, I built a complete machine learning pipeline to detect killfeed elements (kill_icon) in Valorant clips.

From dataset collection to model deployment, I created a YOLOv8-powered system that parses gameplay footage into searchable, taggable, and future-clippable metadata - forming the foundation of an automated content analysis tool.

🔧 What I Built

• 💾 400+ frame training dataset, all custom labeled
• 🎯 YOLOv8 object detector, trained from scratch
• 🧪 v0 → v5 model iterations with real improvement tracking
• 📊 95.96% accuracy on unseen test clips
• 🧠 Full data pipeline using Label Studio, OpenCV, Python, and PyTorch
• 📎 Auto-labeling system with feedback loop ready
• 🎥 OBS integration plan for live data generation
• 🚀 Future integration with LLMs and SEEM