Mobilefacenet with Tensorflow-2, EdgeTPU models also supplied for running model on Coral EdgeTPU
Tensorflow 2 version of mobilefacenet from MobileFaceNets: Efficient CNNs for Accurate Real-Time Face Verification on Mobile Devices
Running 60 fps on Desktop with Coral TPU, and around 24 fps on raspberry pi
Use the same dataset as used in Mobilefacenet-Pytorch to train. CASIA is used for training and LFW is used for testing.
Change the directory pointing to image dataset in train.py. I trained the model directly with ArcFace by setting RESUME to False but it is worthwhile to try out pretraining with softmax loss
I added an example to add extra header to perform classification using generated embedding, here I use generated embedding to make prediction on whether a person is wearing mask. You can have more fun by using another dataset
Trained model is evaluate on each epoch use LFW dataset and I got 99.3% accuracy without pretraining