Face detector/embeddings based on MTCNN, tensorflow and golang
Implementation based on https://github.com/davidsandberg/facenet . Tensorflow (1.4.1) and the golang binding are required.
Model file cmd/mtcnn.pb is converted from facenet too (see scripts/convert.py. You will need to add facenet/src to PYTHONPATH to use it). You may need to regenerate the model file for a different version of tensorflow.
The facenet protobuf model file is available for download (see instructions from facenet).
// detection
bs, err := ioutil.ReadFile(*imgFile)
img, err := goface.TensorFromJpeg(bs)
det, err := goface.NewMtcnnDetector("mtcnn.pb")
bbox, err := det.DetectFaces(img) //[][]float32, i.e., [x1,y1,x2,y2],...
// embeddings
mean, std := goface.MeanStd(img)
wimg, err := goface.PrewhitenImage(img, mean, std)
fn, err := goface.NewFacenet("facenet.pb")
emb, err := fn.Embedding(wimg)
See cmd/detect.go. Use go build to build the binary and run with --help.
- Not exactly the same (e.g., nms/padding is depending on tensorflow implementation).
- Not fully tested. Performance could a little bit worse.
- Face landmark support not implemented.