Short answers, pulled from the story.
Florian Schroff, Dmitry Kalenichenko, and James Philbin unveiled FaceNet at the 2015 IEEE Conference on Computer Vision and Pattern Recognition. These three researchers were affiliated with Google when they presented their work on facial recognition.
FaceNet achieved an accuracy score of 99.63 percent on the Labeled Faces in the Wild dataset. This result represented the highest score recorded on LFW using the unrestricted protocol with labeled outside data.
The NN1 model utilized a total of 140 million parameters across its convolutional layers. This architecture required approximately 1.6 billion floating-point operations per second during processing.
A key innovation involved the triplet loss function which mapped face images into a 128-dimensional Euclidean space. Similarity between faces was assessed based on the square of the Euclidean distance between normalized vectors in that space.
FaceNet reached an accuracy of 95.12 percent when tested against the YouTube Faces DB dataset and 99.63 percent on the Labeled Faces in the Wild dataset. These figures demonstrated superior performance compared to existing methods at the time of publication.