I based my work on FaceNet, a quite recent paper that achieves state-of-the-art results and incredible robustness using very similar techniques to the ones I implemented here (siamese networks, contrastive/triplet loss).