YuNet 2023mar (ONNX, dynamic input size)
Face detector used by redact-web to find faces to redact, in the browser.
- Upstream: https://github.com/opencv/opencv_zoo/tree/main/models/face_detection_yunet (
face_detection_yunet_2023mar.onnx), by Shiqi Yu and contributors. - License: MIT, the same as upstream.
- Training data: WIDER FACE, which is licensed for research use only. The weights license (MIT) is what governs their use; the training data's terms are disclosed here and in redact-web's NOTICE.
Our modifications
Built by tools/models/yunet/export.py:
- The input height and width are made dynamic (
H,W) and the output anchor count dynamic (N). The graph is fully convolutional, so the weights are unchanged. - The export checks the dynamic graph against the original at 640x640.
Files
| File | Description |
|---|---|
onnx/model.onnx |
fp32 graph, input input float32 [1, 3, H, W] (BGR, 0 to 255) |
redact-web pins this repository at a commit and verifies every file's SHA-256 before use.