Teaching a computer to watch hockey.
A machine learning system that detects when a goal is scored on a live broadcast and triggers lights behind my TV. Started with the Leafs, being built to work for any team.
I wanted my living room to react when the Leafs score. Here's the full system design and why I chose computer vision over an API.
Before you can train a model you need data. Before you have data you need footage. This is how I got it and why highlight reels beat full broadcasts for this problem.
60 videos. 210,000 frames. About 0.5% of them matter. This is how I got the frames out — and why I'm only saving a 400×100 pixel crop of each one.