Labeled Faces in the Wild (LFW)
University of Massachusetts Amherst
13,000 news photos of 5,700 public figures, gathered without consent, that set face-recognition benchmarks.
Face photos of public figures collected from news websites, labeled by name, with standard pairs for testing whether two photos show the same person. It was the main face-verification benchmark of the 2010s. It is also a case study in face-data ethics: subjects never consented, it skews toward white men, and larger successors such as Microsoft's MS-Celeb-1M were withdrawn in 2019 after criticism.
Access
- Cost
- FreeFree download.
- License
- License not statedNot formally licensed; photos remain under news publishers' copyrights
- Sign-up
- No sign-up
- How you get it
- Bulk downloadJPEG, TXT pair lists
What's in it
- One row is
- One face image labeled with the person's name
- Covers
- Public figures in early-2000s news photos
- Years
- Released 2007
- Updated
- One-time release
- Size
- 13,233 images of 5,749 people; 1,680 people have two or more images
Quality
Historically central and well documented, but saturated and demographically skewed; the maintainers themselves caution against using it to certify commercial systems.
- Subjects did not consent to inclusion
- Heavily skewed toward white, male, adult faces
- Related face sets such as MS-Celeb-1M and MegaFace were withdrawn over consent concerns
Connects to
- ImageNet: Another web-scraped vision set whose people categories raised similar consent issues
- Wikidata JSON Dumps: Match LFW names to Wikidata to profile who is represented by gender, nationality, and occupation
Questions it could answer
- What is the gender and age breakdown of the people in LFW?
- How do face-verification error rates differ across demographic groups on LFW pairs?