Level Playing Field for Million Scale Face Recognition

摘要

Face recognition has the perception of a solved problem, however when testedat the million-scale exhibits dramatic variation in accuracies across thedifferent algorithms. Are the algorithms very different? Is access to good/bigtraining data their secret weapon? Where should face recognition improve? Toaddress those questions, we created a benchmark, MF2, that requires allalgorithms to be trained on same data, and tested at the million scale. MF2 isa public large-scale set with 672K identities and 4.7M photos created with thegoal to level playing field for large scale face recognition. We contrast ourresults with findings from the other two large-scale benchmarks MegaFaceChallenge and MS-Celebs-1M where groups were allowed to train on anyprivate/public/big/small set. Some key discoveries: 1) algorithms, trained onMF2, were able to achieve state of the art and comparable results to algorithmstrained on massive private sets, 2) some outperformed themselves once trainedon MF2, 3) invariance to aging suffers from low accuracies as in MegaFace,identifying the need for larger age variations possibly within identities oradjustment of algorithms in future testings.