Face Recognition: From Scratch To Hatch
summary
Mail.ru Group builds a production face system: detector + recognizer with 5-landmark alignment, trained on MS‑Celeb plus a cleaned VK crawl (reduced to ~46k persons). They compare Viola‑Jones (45%) vs Faster‑R‑CNN (~92%) and favor an MTCNN cascade for speed; triplet loss underperforms (≈65%) while center loss and angular softmax variants give much better accuracy. Engineering wins: TensorRT + C++ reimplementation cuts inference from ~17ms to ~8–11ms; a two‑model ensemble reaches about 79.6% on MegaFace.
AI-written summary · about
chapters (5)
- 0:00Introduction
- 12:51Metric learning
- 27:19Model errors and improvements
- 38:18Optimization and conclusions
- 40:34Q&A
related talks
40m"Introduction to Face Processing with Computer Vision|"
01h 12mBuilding a live face recognition system
57mFoundations of Facial Recognition
39mBuilding Highly Scalable Facial Recognition Pipelines
31mFace Recognition
26mNew Face, Who Dis? Recent Adversarial Approaches to Facial Recognition
11mSaFace: Towards Scenario-aware Face Recognition via Edge Computing System
33mReal-Time Face Recognition: Training Fast and Accurate Models