In order to extract faces, we use opensource machine learning framework – OpenFace. OpenFace utilizes deep neural networks for face recognition and represents a face of an individual as 128-dimensional hypersphere. The comparision is done by simple difference between two hyperspheres.
All of the portraits are preprocessed and a model of a face extracted from the portrait is cached in order to speed up the computation. The search for the most similar faces is done within a second and the similarity computation takes about 50ms in an average on more than 1 200 extracted faces.