Skip to content

SOURCE.md

Latest commit

 

History

History
67 lines (44 loc) · 3.02 KB

SOURCE.md

File metadata and controls

67 lines (44 loc) · 3.02 KB
Raw

Source and verification

Where every model comes from and how to reproduce or confirm it, so nothing has to be taken on trust.

v2 models (current)

The two v2 files are ONNX exports produced from the upstream published weights. The export scripts (convert_yolov5face.py, convert_ghostfacenet.py) are in convert/.

yolov5face.onnx

  • Upstream: deepcam-cn/yolov5-face (GPL-3.0), the yolov5s-face weights (.pt), from that repository's README download links.
  • Export: clone the repo, then python convert_yolov5face.py weights/yolov5s-face.pt yolov5face.onnx. The script switches the detection head to the concatenated-decoded output and exports at 640x640, opset 12.
  • I/O: input input [1,3,640,640] (RGB, NCHW, /255, letterboxed); one output output [1,25200,16], each row cx,cy,w,h, obj, 5x(x,y) landmarks, class, in 640-letterbox pixels.

ghostfacenet.onnx

  • Upstream: HamadYA/GhostFaceNets (MIT), the GhostFaceNetV1 W1.3 basic-model weights GN_W1.3_S2_ArcFace_epoch48.h5, from that repository's releases.
  • Export: clone the repo, then python convert_ghostfacenet.py GN_W1.3_S2_ArcFace_epoch48.h5 ghostfacenet.onnx (Keras load plus tf2onnx, opset 13).
  • I/O: input input [1,112,112,3] (RGB, NHWC, (x-127.5)/128); output [1,512] embedding (the app L2-normalizes it).

Verify

sha256sum yolov5face.onnx ghostfacenet.onnx

Expected:

8ece145c7a956ed276250778bdb89e40c8ac9521c8669b14d79718cc83ccab32  yolov5face.onnx
ffd8203a0c9e93d90a4957e24d173a883a5a04d64d23c771742c66273518a0db  ghostfacenet.onnx

Sizes: yolov5face.onnx 32370314 bytes, ghostfacenet.onnx 16190333 bytes.

v1 models (previous)

The two v1 files are the original OpenCV Zoo releases, unmodified.

Upstream: OpenCV Zoo, https://github.com/opencv/opencv_zoo (each model carries its own license, noted in NOTICE).

File here Upstream file Upstream directory
yunet.onnx face_detection_yunet_2023mar.onnx models/face_detection_yunet
sface.onnx face_recognition_sface_2021dec.onnx models/face_recognition_sface

The blobs are stored with Git LFS upstream, so use the media host (a plain raw URL returns the LFS pointer, not the file):

curl -L -o yunet.onnx \
  https://media.githubusercontent.com/media/opencv/opencv_zoo/main/models/face_detection_yunet/face_detection_yunet_2023mar.onnx

curl -L -o sface.onnx \
  https://media.githubusercontent.com/media/opencv/opencv_zoo/main/models/face_recognition_sface/face_recognition_sface_2021dec.onnx

Expected:

8f2383e4dd3cfbb4553ea8718107fc0423210dc964f9f4280604804ed2552fa4  yunet.onnx
0ba9fbfa01b5270c96627c4ef784da859931e02f04419c829e83484087c34e79  sface.onnx

Sizes: yunet.onnx 232589 bytes, sface.onnx 38696353 bytes.

  • yunet.onnx: input input [1,3,640,640]; outputs cls_{8,16,32}, obj_{8,16,32}, bbox_{8,16,32}, kps_{8,16,32} per stride.
  • sface.onnx: input data [1,3,112,112]; output fc1 [1,128].