mAP ≥ 70% on event detection, across 18 hours of game video
Models that watch a hockey broadcast and work out what is happening, built for AttentiveVision, a machine-vision startup.
Context
AttentiveVision, a startup co-founded by York professor James Elder, set out to give amateur hockey a broadcast-style viewing experience by following the play automatically.
What I did
I trained a CNN-LSTM stack for event detection on broadcast video, built real-time player tracking with a Kalman filter and OpenPose, and ran an active-learning annotation loop for goals, timeouts and penalties.
Outcome
mAP ≥ 70% on event detection across 18 hours of game video. Tracking under 100 ms at 30 FPS. Labels reached Cohen’s kappa ≥ 0.92.
Evidence
- Vector Institute: Amateur hockey given professional viewing experience courtesy of machine vision startup co-founded by Vector researcher ()
The article describes the project. It doesn’t name individual contributors.
Stack
Python, PyTorch, TensorFlow, OpenCV, OpenPose, CNN-LSTM, Kalman filter