librefacerec-det
Lightweight face detector with 5 facial landmarks: the default upstream
detector for LibreYOLO's facial-recognition (embed) task. Detected boxes
and landmarks feed the canonical 112x112 similarity alignment before
embedding. Auto-downloaded on first use; any LibreYOLO detector or custom
callable can be substituted via face_detector=.
Provenance and license
- Mirrored unmodified from the OpenCV Zoo YuNet release
(
face_detection_yunet_2023mar.onnx), renamed tolibrefacerec-det.onnx. - Weights license: MIT (LICENSE shipped alongside the model in opencv_zoo, copyright Shiqi Yu).
- SHA-256:
8f2383e4dd3cfbb4553ea8718107fc0423210dc964f9f4280604804ed2552fa4(verified identical to upstream at mirror time, 2026-07-28). - Trained by its authors on the WIDER FACE dataset; see the upstream repository for details.
Notes
Runs via OpenCV's bundled cv2.FaceDetectorYN (no extra dependency in
LibreYOLO). Measured 2026-07-28 in the LibreYOLO pipeline on the 2000
funneled LFW dev-test images: a face was found in every image, so no pair
was lost to detection failure. LFW is a frontal, well-lit distribution;
expect lower recall on profile views, small faces, and difficult lighting.