FaceNet

FaceNet is a face recognition system developed by researchers at Google and introduced in 2015. It presents a unified approach to face recognition and verification by learning a compact embedding for facial images. The core idea is to train a deep convolutional neural network to map each facial image to a 128-dimensional vector in Euclidean space, where the distance between embeddings corresponds to facial similarity.

The network is trained with a triplet loss: for a triplet consisting of an anchor, a positive example of the same identity, and a negative example of a different identity, the embedding of the anchor should be closer to the positive than to the negative by a chosen margin. By optimizing on a large labeled dataset, FaceNet learns that images of the same person cluster together while different people are farther apart. After training, embeddings are typically L2-normalized for convenience in distance-based verification and identification.

FaceNet uses a deep CNN architecture, often based on Inception-style backbones, to extract robust features from aligned face crops. The resulting 128-dimensional embeddings can be used for various tasks, including face verification (determining if two images depict the same person), identification within a gallery, and clustering of face images.

FaceNet achieved strong performance on standard benchmarks of its time, such as Labeled Faces in the Wild and YouTube Faces, and its approach has influenced many subsequent face recognition systems. It popularized the use of learned embeddings and metric learning for face recognition and spurred further research into data collection, training techniques, and embedding post-processing.