This is a TensorFlow implementation of the face recognizer described in the paper "FaceNet: A Unified Embedding for Face Recognition and Clustering" http://arxiv.org/abs/1503.03832
The code is heavly inspired by the OpenFace implementation at https://github.com/cmusatyalab/openface
The FaceScrub dataset (http://vintage.winklerbros.net/facescrub.html) and the CASIA-WebFace dataset (http://www.cbsr.ia.ac.cn/english/CASIA-WebFace-Database.html) have been used for training.
The data has been pre-processed as described on the OpenFace web page (https://cmusatyalab.github.io/openface/models-and-accuracies/), i.e. using ./util/align-dlib.py data/lfw/raw align outerEyesAndNose data/lfw/dlib-affine-sz:96 --size 96 --fallbackLfw data/lfw/deepfunneled
A model that has been train on a combination of FaceScrub and CASIA-Webface is available: model-20160506.ckpt-500000 This model has been trained for 500 epochs (with a batch size of 90 images). To load the model Tensorflow needs a checkpoint file in the same directory as the model file. The checkpoint file is created when the model is stored (during training), but can also be created with a text editor (see below). But remember to adjust the paths to point to your model file. checkpoint:
model_checkpoint_path: "/home/david/models/facenet/model-20160506.ckpt-500000"
all_model_checkpoint_paths: "/home/david/models/facenet/model-20160506.ckpt-500000"
The accuracy on LFW for the model "model-20160506.ckpt-500000" is 0.919+-0.008. The test can be run using "validate_on_lfw.py".
This project is currently very much under development (i will try to keep the issue tracker up-to-date with what is in the pipe).