subhashhhhhh/lumen-faces

Face detection and recognition for Lumen

On-device face grouping for Lumen, a local-first Android gallery: it finds the faces in photos and groups them by person. Lumen downloads these files on request from its Settings; no photo or face ever leaves the phone.

Both are unchanged copies from OpenCV Zoo, commit 48b10fa505aeb51fbd127fc60a5c660040e823b7:

File	What	Source	License
yunet.onnx	Face detection: boxes, 5 landmarks (eyes, nose tip, mouth corners), scores. Any input size	face_detection_yunet_2026may.onnx (YuNet, Shiqi Yu)	MIT (LICENSE-YuNet)
sface.onnx	Face recognition: an aligned 112×112 face → 128-d embedding	face_recognition_sface_2021dec.onnx (SFace, Zhong et al. 2021)	Apache 2.0 (LICENSE)

SHA256SUMS lists each file's checksum; Lumen verifies them after download.

Using them

• YuNet: BGR, raw 0–255 values, NCHW, height and width padded to multiples of 32. Outputs per stride (8, 16, 32): cls, obj, bbox, kps; score = √(cls × obj); box centre (col + dx, row + dy) × stride, size exp(dw, dh) × stride; landmarks (col + kx, row + ky) × stride.
• Alignment: a similarity transform (rotation, uniform scale, shift) fitted by least squares from the 5 landmarks to the standard 112×112 template (38.29, 51.70), (73.53, 51.50), (56.03, 71.74), (41.55, 92.37), (70.73, 92.20).
• SFace: the aligned face, RGB, raw 0–255 values, NCHW → 128 values; L2-normalize, compare by cosine similarity. Lumen groups faces at ≥ 0.55.

Training data

OpenCV doesn't state what SFace was trained on (opencv_zoo#313); the SFace paper used research face datasets (CASIA-WebFace, VGGFace2, MS1M). The license above is OpenCV's grant for these files. YuNet was trained on WIDER FACE.

License

YuNet: MIT. SFace: Apache License 2.0. See LICENSE-YuNet, LICENSE and NOTICE.