Large data sets overwhelm facial recognition software

IEEE Spectrum : Many facial recognition algorithms have a success rate above 95% when tested against databases of just a few thousand faces. A new test called the MegaFace Challenge evaluates the performance of algorithms when they are presented with a database of 1 million images of 690 000 people. The challenge is for the algorithms to evaluate whether two different pictures are of the same person and to determine if a given person is in the database. The success rates for all the tested facial recognition programs dropped significantly when faced with so much data. Google’s FaceNet, the top-scoring algorithm tested, dropped from near 100% accuracy on the widely used Labeled Faces in the Wild (LFW) test to just 75% on the MegaFace Challenge. Several algorithms that scored above 90% on LFW dropped to below 60% accuracy. Ira Kemelmacher-Shlizerman of the University of Washington in Seattle and his colleagues organized the MegaFace Challenge to evaluate the effectiveness of facial recognition software in more realistic situations.