Cabbage, or recognizing 330 million faces at 1000 photos/sec
summary
Describes a high-throughput face-recognition system for a social network with 330 million user accounts and 20 million photos uploaded daily. Presents an approach that processes each photo in under 0.2 seconds using limited hardware resources to achieve recognition at about 1,000 images per second.
AI-written summary · about
chapters (20)
- 0:00Introduction
- 1:30Face recognition challenge
- 3:39What are neural networks
- 5:46Convolutional neural networks
- 7:53Recognition pipeline
- 8:55Face detector optimization
- 12:10Detector accuracy improvements
- 14:24Cosine similarity and vectors
- 15:47Training neural networks from scratch
- 19:20Loss function and training
- 22:34Face search optimization
- 24:20False positives problem
- 25:45User database search
- 27:51Caching and optimization
- 29:58Testing and tuning
- 31:45User profile building
- 34:36System scaling
- 36:47Final pipeline
- 38:57Museum Night application
- 40:24Q&A
related talks
31mFace Recognition
40m"Introduction to Face Processing with Computer Vision|"
01h 12mBuilding a live face recognition system
32mFace Pay
21mLeveraging Python for Real-Time Image processing
57mFoundations of Facial Recognition
30mIn the service of the history. AI in archivistics.
50mDistributed Facial Recon