# FaceNet: A Unified Embedding for Face Recognition and Clustering

**Authors:** Florian Schroff, Dmitry Kalenichenko, James Philbin  
**arXiv:** [1503.03832](https://arxiv.org/abs/1503.03832)  
**Date:** 2015-03-12  
**Summary page:** https://www.summarizepaper.com/en/paper/1503.03832v3/

## Key points

- Face recognition and verification at scale are challenging tasks for current approaches
- Authors present a system called FaceNet that directly learns a mapping from face images to a compact Euclidean space where distances correspond to a measure of face similarity
- This approach enables efficient implementation of tasks such as face recognition, verification, and clustering using standard techniques with FaceNet embeddings as feature vectors
- The authors' method uses a deep convolutional network trained to optimize the embedding itself, unlike previous deep learning approaches that use an intermediate bottleneck layer
- To train the model, they use triplets of roughly aligned matching/non-matching face patches generated using an online triplet mining method
- Their system achieves state-of-the-art face recognition performance using only 128-bytes per face on Labeled Faces in the Wild (LFW) dataset and YouTube Faces DB.
- On LFW dataset, their system achieves a new record accuracy of 99.63%, while on YouTube Faces DB it achieves 95.12%
- Compared to the best published results on both datasets, their system cuts down the error rate by 30%
- The authors also introduce the concept of harmonic embeddings and harmonic triplet loss that describe different versions of face embeddings produced by different networks that are compatible with each other and allow for direct comparison between them.
- Overall, FaceNet's direct learning approach offers significant improvements in representational efficiency and accuracy compared to previous methods for implementing face recognition and verification at scale.

## Summary

Face recognition and verification at scale have been challenging tasks for current approaches, despite recent advances in the field. In this paper titled "FaceNet: A Unified Embedding for Face Recognition and Clustering," authors Florian Schroff, Dmitry Kalenichenko, and James Philbin present a system called FaceNet that directly learns a mapping from face images to a compact Euclidean space where distances correspond to a measure of face similarity. This approach enables efficient implementation of tasks such as face recognition, verification, and clustering using standard techniques with FaceNet embeddings as feature vectors. Unlike previous deep learning approaches that use an intermediate bottleneck layer, the authors' method uses a deep convolutional network trained to optimize the embedding itself. To train the model, they use triplets of roughly aligned matching/non-matching face patches generated using an online triplet mining method. The result is much greater representational efficiency, achieving state-of-the-art face recognition performance using only 128-bytes per face. On the widely used Labeled Faces in the Wild (LFW) dataset, their system achieves a new record accuracy of 99.63%, while on YouTube Faces DB it achieves 95.12%. Compared to the best published results on both datasets, their system cuts down the error rate by 30%. The authors also introduce the concept of harmonic embeddings and harmonic triplet loss that describe different versions of face embeddings produced by different networks that are compatible with each other and allow for direct comparison between them. Overall, FaceNet's direct learning approach offers significant improvements in representational efficiency and accuracy compared to previous methods for implementing face recognition and verification at scale.