#opencv

Face Recognition on One CPU Core: I Tested 28 Video Pipelines and the Biggest Model Didn’t Always Win

A hands-on benchmark comparing 28 face recognition video pipelines running on a single CPU core. The author tested 20 recognition model setups (including ArcFace MobileFaceNet, FaceLiVT, SFace, and others) with detectors like SCRFD, YuNet, RetinaFace, and dlib HOG. Key finding: larger models don't always win — three small ArcFace MobileFaceNet pipelines (InsightFace Buffalo-S, SCRFD+ArcFace MBF with ONNX Runtime, UniFace ArcFace MNet) tied for best results at 16/17 clip recall with no false identities and ~12.7–12.8 Corpus FPS. OpenCV YuNet+SFace is highlighted as a strong alternative with clear commercial licensing. The post also covers methodology, metrics definitions, dataset limitations, and license status of tested models.