Engineering Lab Facial Recognition

Recognize different people's faces and greet them verbally with just a single picture of each person. Works well with bad lighting and non-frontal poses.

Setup

• Install python3
• Install git-lfs
• Download pretrained models and put them in the saved_models top level directory
• Install mpg321 with a package manager or here
• Run

git clone --recursive https://github.com/StPauls-Computer-Science/ng-lab-face-recognition.git
pip install -r requirements.txt

Running Recognition

Find a single picture of everyone you want to recognize and put them in a folder. Then, generate a database of embeddings with

python create_database.py --use-fixed-standardization (output_file) \
    (input_directory)

Then, you can do real-time camera facial recognition by running

python camera_recognize.py (database_path)

where database_path refers to the file you just generated. The filenames will be used as labels (people's names).

Notebook

In the notebook Facial Recognition with FaceNet.ipynb, I've been exploring the embeddings and building a simple face recognizer. Check it out for a brief explanation of FaceNet and the embeddings.

A Note About Test Images

I've provided three pictures of myself and Sam Henderson, who agreed to have his pictures put online. For other pictures referenced in the notebook, get some pictures from Google Images.

Visualizing the saved model

python log_saved.py
tensorboard --logdir logs/facenet-pretrained-log

Will give you cool graph visualizations in tensorboard so you can explore the model visually.

Acknowledgements

I'm using David Sandberg's implementation of FaceNet (included as a submodule) and his pretrained weights. I'm also using OpenFace's wrapper around dlib's face detection and alignment. OpenFace is an open source face recognition module developed by Carnegie Mellon. To find out more about FaceNet, read the original paper on arxiv here