How to use facenet: clone, set up and run the face recognition code

What facenet is and what it includes

facenet is a TensorFlow implementation of the face recognizer described in the paper "FaceNet: A Unified Embedding for Face Recognition and Clustering". The project also uses ideas from the paper "Deep Face Recognition" from the Visual Geometry Group at Oxford, and its code is heavily inspired by the OpenFace implementation.

The repository includes training code, face alignment using MTCNN, and pretrained models. The README reports that the best pretrained model reaches an LFW accuracy of 0.99650+-0.00252.

• Training code for models using softmax loss and triplet loss
• Face alignment with a Python/TensorFlow implementation of MTCNN in src/align
• Pretrained Inception-ResNet-v1 models trained on CASIA-WebFace and VGGFace2

Check the compatibility before you start

The README is explicit about what the code was tested with. It says the code is tested using TensorFlow r1.7 under Ubuntu 14.04 with Python 2.7 and Python 3.5. Older news entries also say it was updated to run with TensorFlow r1.0 and r0.12, and that it is not sure whether it runs with older versions.

If you plan to use a current TensorFlow or Python release, expect to do compatibility work yourself. The README does not say the code works with newer versions.

• Tested: TensorFlow r1.7, Ubuntu 14.04, Python 2.7 and 3.5
• Not documented: current TensorFlow or Python releases

Clone the repository

The install method is clone-and-build. Start by cloning the repository from GitHub with the command in the README. The project is not packaged as a pip install, so you work from the cloned source folder.

Choose a pretrained model

The README lists two pretrained models. Both use the Inception ResNet v1 architecture. Model 20180408-102900 was trained on CASIA-WebFace with an LFW accuracy of 0.9905. Model 20180402-114759 was trained on VGGFace2 with an LFW accuracy of 0.9965, and the README describes VGGFace2 as the dataset behind the best performing model.

The models need fixed image standardization. The README says to add the option --use_fixed_image_standardization when running a script such as validate_on_lfw.py. If you use any of the models, the README asks you to credit the providers of the training dataset.

• Use --use_fixed_image_standardization with validation scripts such as validate_on_lfw.py
• Credit the training dataset providers when you use a model

Validate on LFW and train your own model

To check accuracy, the README points to the wiki page Validate on lfw. For training, it says the best results come from softmax loss. The README links to a wiki page on classifier training of Inception-ResNet-v1 on CASIA-WebFace, and a separate wiki page on training a classifier on your own images.

Note that the README's news section says the older facenet_train.py was renamed to train_tripletloss.py and facenet_train_classifier.py was renamed to train_softmax.py. Use the current file names when you follow older tutorials.

• Validation guide: the wiki page Validate on lfw
• Softmax training guide: Classifier training of Inception-ResNet-v1
• Own-image classifier guide: Train a classifier on own images

Understand what the alignment step does

The README explains that face alignment uses MTCNN, a multi-task CNN. The author found that the Dlib face detector missed some hard examples such as partial occlusion and silhouettes, which made training data too easy. The repository's Python/TensorFlow MTCNN implementation is in src/align. The README says it does not give identical results to the Matlab/Caffe version, but performance is very similar.