Lightweight face matching with OpenCV YuNet and SFace

To check whether a known person appears across about 100 photos, cv2.FaceDetectorYN (YuNet onnx, ~230KB) and cv2.FaceRecognizerSF (SFace onnx, ~38MB) from opencv_zoo need only opencv-python, with no dlib or face_recognition install. Reference embeddings can come from a single collage containing several faces. Cosine thresholds that worked: best score >=0.40 is a match, 0.30-0.40 is unsure, below 0.30 is no match. Misses clustered on upside-down faces, tiny faces, scans of old prints and much younger versions of the person, so low scores were mostly false negatives. Build a contact sheet of the low scorers and have a human confirm them rather than auto-rejecting.