foundyou-models

The two ONNX graphs FoundYou runs on-device to group photos by the people in them. Mirrored here for distribution; neither is trained by this project.

file	role	bytes	licence
face_detection_yunet_2023mar.onnx	detector	2,32,589	Apache-2.0
face_recognition_sface_2021dec.onnx	embedder	3,86,96,353	Apache-2.0

Both are published by OpenCV Zoo and are Apache-2.0 including their weights, which is why they can ship in a commercial app. They are byte-identical to the upstream files — no conversion, no quantisation, no re-export.

Verify before you use these

8f2383e4dd3cfbb4553ea8718107fc0423210dc964f9f4280604804ed2552fa4  face_detection_yunet_2023mar.onnx
0ba9fbfa01b5270c96627c4ef784da859931e02f04419c829e83484087c34e79  face_recognition_sface_2021dec.onnx

FoundYou checks every download against these hashes, compiled into the app binary rather than read from this repository. A mirror serving different bytes is rejected rather than trusted, which is what makes public hosting acceptable for executable content.

Models

face_detection_yunet_2023mar.onnx

YuNet (2023mar). 640x640 NCHW input named input; twelve outputs — cls_N, obj_N, bbox_N, kps_N for strides 8, 16 and 32. Pixels are BGR, unscaled 0-255.

Two decoding details produce plausible-looking output rather than an error when got wrong: the per-cell score is the geometric mean of the classification and objectness heads, and the box extents are logarithmic (exp(v) * stride).

It predicts all five landmarks an ArcFace-family embedder aligns against — both eyes, nose tip, and both mouth corners. Detectors that emit a mouth centre instead force two of the five to be synthesised, which degrades every embedding produced downstream.

face_recognition_sface_2021dec.onnx

SFace, a MobileFaceNet trained with the SFace loss. 112x112 NCHW input named data, output fc1 of 128 floats, L2-normalised by the caller.

Pixels are raw 0-255 RGB — not (x - 127.5) / 128. Measured against OpenCV's reference implementation over 100 photographs, raw RGB agrees at cosine 1.000000 and the ArcFace-style normalisation at 0.93. A 0.93 agreement still "works": faces are found, embeddings look reasonable, and every comparison is quietly worse.

Licence

Apache-2.0, matching upstream. This repository adds no weights of its own, so anyone may use these files under Apache-2.0 terms.

That applies to these model files only. FoundYou itself — the application this repository serves — is proprietary and all rights are reserved; mirroring its models here grants no right to its source.