How face recognition works: detection to embeddings

9 steps · about 4 h 53 min · Intermediate

  1. What is Face Embedding? (reference, Mixpeek, 8 min)

    It pins down the key terms, including 5-point alignment, 112x112 crops, 512-d vectors and L2 normalisation.

  2. FaceNet: A Unified Embedding for Face Recognition and Clustering (doc, arXiv (Schroff, Kalenichenko, Philbin), 45 min)

    It is the original paper that framed faces as embeddings where distance means similarity.

  3. ArcFace: Additive Angular Margin Loss for Deep Face Recognition (doc, arXiv (Deng et al.), 1 h)

    It is the original paper for the loss behind most modern face models, so read it once you know what an embedding is.

  4. serengil/retinaface: RetinaFace: Deep Face Detection Library for Python (repo, GitHub (Sefik Serengil), 20 min)

    It lets you try the detection and landmark stage on its own and see the face crops it produces.

  5. serengil/deepface: A Lightweight Face Recognition and Facial Attribute Analysis Library for Python (repo, GitHub (Sefik Serengil), 40 min)

    It gives you a free Python pipeline to verify two faces and compare models such as FaceNet and ArcFace.

  6. How to build a Face Recognition System with Endee (tutorial, Endee, 30 min)

    It scales one-to-one matching up to searching a database of stored embeddings, which is how real systems identify people.

  7. Face Recognition Performance, Bias, and the Limits of Technical Fixes (article, Federation of American Scientists, 20 min)

    It shows that accuracy differs across groups, which you need to know before trusting any headline accuracy number.

  8. NISTIR 8280 Face Recognition Vendor Test (FRVT) Part 3: Demographic Effects (doc, NIST, 45 min)

    It is the primary source on how false match and false non-match rates are measured across demographics.

  9. nataphong-lab/Open-Source-Face-Recognition-Ultimate (repo, GitHub (nataphong-lab), 25 min)

    You can now read this on-device SDK and its benchmark claims critically, asking what was tested and on which datasets.