How the NFL builds computer vision training datasets at scale
summary
NFL video gets time-synced multi-view preprocessed, then labeled with Amazon SageMaker Ground Truth using vendor crowds plus private experts, assistive AI auto-tracks helmets and a consensus algorithm consolidates answers. Labels use per-play jersey-based IDs and V00/H00 tags for unknowns to prevent ID switching; audit checks flag sudden bbox jumps for relabeling. Dataset covers 48 reviewed games, ~22 players per play and iterative rounds feed models that track impacts; project claims 30% concussion reduction and runs a $75,000 Kaggle challenge.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 0:52Machine Learning Lifecycle
- 6:08Data labeling workflows
- 7:48Human workforce options
- 9:45Sam Huddleston use case
- 13:08Identifying definitive helmet impacts
- 14:31Labeling procedures
- 17:27Key lessons learned
- 18:51Improved labeling scheme
- 21:38Tags to add context
- 23:20Managing at scale
- 27:11Tracking
- 28:06Competition
related talks
33mUsing machine learning to improve player safety in the NFL
42mHow Deep Learning is Revolutionizing Sports Analytics
51mVideo Analytics for Football games
27mAnalysis of sports scenes using computer vision models
31mBeyond a Bit Fit
28mScalable High Precision OCRs for Game Event Detections
35mFootball video analysis using Deep Learning
36m"Lights, camera, AI!" - Automated sports videography