Deep Learning for Automatic Basketball Video Production
summary
A system that turns a fixed wide camera into a live, watchable basketball broadcast by combining Gaussian-based actionness (mixture of player-position Gaussians projected to 1D) with learned game-state classifiers and a modified YOLOv3 detector tuned to spot the small ball. It outputs discrete virtual framings with PID smoothing, runs in real time adding about 300–400 ms latency, and works well in most games though failures occur when ball detection or lighting breaks down.
AI-written summary · about
chapters (13)
- 0:00Introduction and background
- 3:14Genius Sports overview
- 8:41Automatic basketball production goals
- 11:54Paper contributions explained
- 14:24Modular system vs end-to-end
- 16:50Basketball production styles
- 19:40Action heatmap and Gaussian modeling
- 23:15Actor detection and localization
- 27:12Real-time processing pipeline
- 31:07Game state and production rules
- 35:46Future improvements and research
- 38:57Generalizing to other sports
- 40:18Future directions and vision
related talks
35mFootball video analysis using Deep Learning
20mAI-Driven Real-Time Ball Tracking for Live Sports Streaming
29mAI in Sports - Basketball Rule Violation Detection
27mAnalysis of sports scenes using computer vision models
51mVideo Analytics for Football games
36m"Lights, camera, AI!" - Automated sports videography
38mDeep Learning plays Handball
41mThe evolution of basketball with data science