Testing Live Video Streams At Scale
summary
Company runs 500+ live streams and builds automated visual tests using OpenCV and neural nets to catch real failures. Exposure test masks card whites and averages brightness across ~1,000 frames to decide pass/fail. LED detection moves from Haar cascades to a VGG19-based regressor giving ~3-pixel precision and ~98% presence accuracy, failing if >14px. Histogram detects dusty lenses, change detection uses seven-hour averages, and clothes/shoe-wear NN reaches about 97% shoe detection with MAE ≈0.4; production data forces retraining.
AI-written summary · about
chapters (10)
- 0:00Introduction to stream QC
- 1:15Common visual defects
- 3:52Trouble Buster architecture
- 5:26Exposure and image analysis
- 8:34LED and machine learning
- 11:01Developing the detection model
- 18:54Histogram based detection
- 20:35Change detection and issues
- 26:22Monitoring for equipment wear
- 31:36Conclusion and Q&A
related talks
41mPractical Aspects of Applying Modern Computer Vision Technologies
38mHow We Built a Safety Violations Detection System
38mHow to Test a learning system
01h 12mBuilding a live face recognition system
10mThe Colorful Truth of Automated Tests
33mRunning Machine Learning in Production - a Journey to Success
14mComputer vision on edge devices
45mVisual Regression Testing with Neural Networks