Face Recognition Suite (commercial-friendly)
ONNX models used by the Shoot2Drive Android app and its web gallery viewer for on-device face detection and face search. Every model here is distributed under the Apache License 2.0.
| File | Size | Source | License | Purpose |
|---|---|---|---|---|
blazeface_front.onnx |
440,781 B | MediaPipe BlazeFace (short range), ONNX conversion from PINTO0309 Model Zoo (030_BlazeFace) |
Apache-2.0 | Face detection, 1x128x128x3 RGB NHWC, (x-127.5)/127.5 |
blazeface_back.onnx |
603,172 B | MediaPipe BlazeFace (back camera / long range), ONNX conversion from PINTO0309 Model Zoo | Apache-2.0 | Face detection, 1x256x256x3 RGB NHWC, (x-127.5)/127.5 |
sface.onnx |
38,696,353 B | OpenCV Zoo SFace, face_recognition_sface_2021dec.onnx (renamed, bytes unchanged) |
Apache-2.0 | Face embedding, 128-d |
sface.onnx
- Upstream: https://github.com/opencv/opencv_zoo/tree/main/models/face_recognition_sface
- Pinned file: https://github.com/opencv/opencv_zoo/raw/ba91a3b91d00d76e86540d4013f944bd6b514e39/models/face_recognition_sface/face_recognition_sface_2021dec.onnx
- SHA256:
0ba9fbfa01b5270c96627c4ef784da859931e02f04419c829e83484087c34e79 - Paper: Zhong et al., "SFace: Sigmoid-Constrained Hypersphere Loss for Robust Face Recognition", IEEE TIP 2021.
- Input
data: float32[1,3,112,112], RGB, NCHW, raw pixel values 0..255 (the graph normalizes internally; do not pre-normalize). The face is aligned to the standard ArcFace 112x112 template. - Output
fc1: float32[1,128]. L2-normalize; cosine similarity is the dot product. - Measured with the Shoot2Drive BlazeFace + 4-point alignment pipeline on LFW (6000 pairs, 10-fold): 99.45% Β± 0.32.
- Cosine thresholds (LFW all cross-identity pairs): FAR 1e-3 β 0.364, FAR 1e-4 β 0.421, FAR 1e-5 β 0.472.
Attribution
- SFace model: Copyright the OpenCV Zoo contributors and the SFace authors, Apache License 2.0.
- BlazeFace models: Copyright Google LLC (MediaPipe), Apache License 2.0; ONNX conversion by Katsuya Hyodo (PINTO0309).
The full license text is in LICENSE.
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support