Sort similar looking face pictures to clusters using FaceNet and KMeans

Prerequisites

• installed dependencies tensorflow and opencv libraries.

pip install tensorflow
conda install opencv

• correct PYTHONPATH=full_path_to_folder/scr
• pretrained FaceNet model in the folder ./models/. You can download it here.
• images dataset in the ./data/folder_with_images/. I used LFW dataset. Unzip it to ./data/lfw/. The directory should be in the correct openface format:

my_database
└───a_person
│   │   image00.jpg
│   │   image01.jpg
│   
└───b_person
│   │   image00.jpg
│   
└───c_person
│   │   image00.jpg
│   │   image01.jpg
│   │   image02.jpg
│   │   image03.jpg

How to run

1. Clean folders ./np_embeddings , ./data/clustered, ./data/sorted

• Run export_embeddings.py
  This will generate embeddings and labels for images.
  To make it working add PYTHONPATH=full_path_to_folder/scr to you sources. If you are using PyCharm simply add to Enviroment Varianbles PYTHONPATH value full_path_to_folder/scr.

• Run Distance_matrix.py.
  This step will give you the matrix with all of the Euclidean distances between faces. Numpy array is saved to ./np_embeddings/embeddings.npy It will take some time because the matrix size of NxN with zeros on the main diagonal. This file is quiet big - 1.4GB. We need it to sort faces.

3. Run Cluster_faces.py.
   First, it will sort all off the face images based on the closest distance and save sorted images to ./data/sorted. Second, it will cluster images using Kmeans algorithm. Number of cluster by default is 30. You can change it if you like.

Result

This is what you get in the end. KMeans does a pretty good job to cluster 128-dimentional image embeddings.

Disclaimer

This work is based on FaceNet achievement. You can check FaceNet model and papers here.