Labeled Faces in the Wild : Updates and New Reporting Procedures

The Labeled Faces in the Wild (LFW) database has spurred significant research in the problem of unconstrained face verification and other related problems. While careful usage guidelines were established in the original technical report describing the database, certain unforeseen issues have arisen. One of the major issues is how to make fair comparisons among algorithms that use additional “outside data”, i.e., data that is not part of LFW, for training. Another issue is the need for a clear definition of the “unsupervised paradigm” and the proper protocols for producing results under this paradigm. This technical report discusses these issues in detail and provides a new description of how we curate results and how we group algorithms together based on the details of the training data that they use. We encourage any authors who intend to publish their results on LFW to read both the original technical report and this one carefully.