Labeled Faces in the Wild (LfW)
Technical information:
| Full name | |
|---|---|
| Country | |
| Contents | |
| Images | 13,872 |
| Individuals | 1,462 |
| Runs database software | |
| URL"URL" is a type and predefined property provided by Semantic MediaWiki to represent URI/URL values. | |
| Related Technology |
Developers and Users:
| Developed by | |
|---|---|
| Owning institution | |
| Custodian institution |
Description[edit | ]
The dataset leveraged the Web to source the first fully unconstrained face dataset with over 13,000 images of 1,680 faces in an infinite combination of poses, illumination conditions, and expressions (Huang et al. 2007). 1
LFW inspired a flurry ofWeb-scraped face datasets for facial recognition model training and benchmarking - including many datasets sourcing images without consent from online platforms, such as Google Image search (Bainbridge, Isola, and Oliva 2013; Han et al. 2017; Cao et al. 2018b), Youtube (Chen et al. 2017; Dantcheva, Chen, and Ross 2012), Flickr (Merler et al. 2019; Kemelmacher-Shlizerman et al. 2016) and Yahoo News (Jain and Learned-Miller 2010). 1
As the appetite for unstructured and unconstrained “in the wild” data grew, there was also in this period a proliferation of benchmarks like ChokePoint (Wong et al. 2011) and SCface (Grgic, Delac, and Grgic 2011), datasets that source face images from mock or real surveillance set ups.1
As datasets began to more closely resemble real world conditions, so did the evaluations of commercial products. The Facial Recognition Vendor Test (FVRT) evolved extensively over this period (Blackburn, Bone, and Phillips 2001; Phillips et al. 2003; Ngan, Ngan, and Grother 2015), growing from 13,872 images of about 1,462 subjects in the initial implementation in 2000 to 30.2 million"million" can not be assigned to a declared number type with value 30.2. still photographs of 14.4 million"million" can not be assigned to a declared number type with value 14.4. individuals in the iteration in 2013.1
The research problem of identifying faces in unconstrained conditions nevertheless remained a stubborn technical challenge and development stalled as academics struggled to develop methods to represent faces independently of a controlled image context and template appearance.1