FaceX 02 / 25
Face recognition engine
MobileFaceNet + ArcFace recognition trained from scratch, packaged for the browser and for native C.
Problem
Off-the-shelf recognition stacks (dlib, FaceNet, InsightFace) are heavy and usually need Python or a GPU server.
What it does
- Computes face embeddings and compares identities
- Four model sizes from 0.8 to 8.4 MB to trade accuracy vs footprint
- Native build (~148 KB engine) and WASM build
How it works
- Detected + aligned face → MobileFaceNet backbone trained with ArcFace loss on MS1M
- Embedding → cosine similarity for verification / search
- Evaluated with the standard 10-fold LFW protocol
Stack & models
CWebAssemblyMobileFaceNetArcFaceMS1MLFW
Numbers
- LFW accuracy improved 95.62% → 99.07% (10-fold mean)
- ~3.0 ms per face (README latency badge)
Figures come from project docs/README; no other benchmarks are claimed.