We used LFW dataset to train our model. Our model uses pretrained inception network to generate embeddings and MTCNN for face extraction. We are getting 96.5% accuracy on the dataset.
For the gui application however we have used pretrained facenet network (to generate embeddings) from keras to get maximum accuracy, next we trained a network to predict whether two embeddings are of same person or not.
The Face extraction file extracts faces from the dataset this may take around 1 hour but is a one time process. The utils function loads the data.
To get the model that performs classification use the below link. We are getting overall accuracy of 99% when using facenet + predict_model.
Link for predict_model: https://drive.google.com/file/d/1Y0WH3jUDJo1AMNw1PedqXFsXBP-EZUbg/view?usp=sharing
Dependencies:
Tensorflow, Keras, MTCNN (Face Extraction), FaceNet (Generate Embeddings), Opencv (Capture Images)
Demo: https://drive.google.com/file/d/1sUF7U9AuEPn6TgfhTJqjFj8j3ILjV4AB/view?usp=sharing
Credits: Ipazc https://github.com/ipazc/mtcnn (MTCNN)
Fausto Morales https://github.com/faustomorales/keras-facenet (FaceNet)
MIT License