FaceNet is a system designed to efficiently implement face verification, recognition, and clustering tasks using a deep convolutional network. The network learns a mapping from face images to a compact Euclidean space where distances directly correspond to face similarity. This approach offers greater representational efficiency, achieving state-of-the-art performance with only 128 bytes per face. FaceNet uses a triplet loss function to train the network, which optimizes the embedding directly, rather than an intermediate bottleneck layer. The method employs a novel online triplet mining technique to select hard triplets for training, ensuring consistent difficulty as the network trains. On the Labeled Faces in the Wild (LFW) dataset, FaceNet achieves a new record accuracy of 99.63%, and on YouTube Faces DB, it achieves 95.12%, reducing the error rate by 30% compared to the best published results. The paper also introduces the concept of harmonic embeddings and a harmonic triplet loss, which allow for the comparison and compatibility of embeddings produced by different networks.FaceNet is a system designed to efficiently implement face verification, recognition, and clustering tasks using a deep convolutional network. The network learns a mapping from face images to a compact Euclidean space where distances directly correspond to face similarity. This approach offers greater representational efficiency, achieving state-of-the-art performance with only 128 bytes per face. FaceNet uses a triplet loss function to train the network, which optimizes the embedding directly, rather than an intermediate bottleneck layer. The method employs a novel online triplet mining technique to select hard triplets for training, ensuring consistent difficulty as the network trains. On the Labeled Faces in the Wild (LFW) dataset, FaceNet achieves a new record accuracy of 99.63%, and on YouTube Faces DB, it achieves 95.12%, reducing the error rate by 30% compared to the best published results. The paper also introduces the concept of harmonic embeddings and a harmonic triplet loss, which allow for the comparison and compatibility of embeddings produced by different networks.