YuNet Face Detection — 2026may (dynamic input)
Unmodified redistribution of face_detection_yunet_2026may.onnx from the
OpenCV Zoo.
YuNet is a light-weight, fast and accurate face detection model (WIDER Face: 0.834 AP_easy / 0.824 AP_medium / 0.708 AP_hard) that detects faces of roughly 10×10 to 300×300 px. This is the dynamic-input-shape variant (the OpenCV Zoo default): compatible with OpenCV 5.x's ONNX Runtime engine and any ONNX runtime, allowing inference at any resolution without resizing. It is the same model with symbolic height/width — the weights are identical to the fixed-shape 2023mar variant.
Input
input — float32 tensor [1, 3, height, width] (dynamic H/W): a BGR image, NCHW
layout, no normalization. height and width must be multiples of 32 (the largest stride).
Output
12 per-stride detection heads for strides s ∈ {8, 16, 32}, each shaped [1, A, C]
with A = (height/s) · (width/s) anchors:
| tensor | shape | meaning |
|---|---|---|
cls_{s} |
[1, A, 1] |
classification score |
obj_{s} |
[1, A, 1] |
objectness score |
bbox_{s} |
[1, A, 4] |
bounding-box regression |
kps_{s} |
[1, A, 10] |
5 landmarks — right eye, left eye, nose, right/left mouth corner — as (x, y) |
cv2.FaceDetectorYN (or an equivalent decode) turns these heads into per-face
[x, y, w, h, 5 × (x, y), score].
License
MIT, following the upstream OpenCV Zoo.
@article{wu2023yunet,
title={YuNet: A Tiny Millisecond-level Face Detector},
author={Wu, Wei and Peng, Hanyang and Yu, Shiqi},
journal={Machine Intelligence Research},
volume={20}, number={5}, pages={656--665}, year={2023}, publisher={Springer}
}