BioMetricNet: deep unconstrained face verification through learning of metrics regularized onto Gaussian distributions

We present BioMetricNet: a novel framework for deep unconstrained face\nverification which learns a regularized metric to compare facial features.\nDifferently from popular methods such as FaceNet, the proposed approach does\nnot impose any specific metric on facial features; instead, it shapes the\ndecision space by learning a latent representation in which matching and\nnon-matching pairs are mapped onto clearly separated and well-behaved target\ndistributions. In particular, the network jointly learns the best feature\nrepresentation, and the best metric that follows the target distributions, to\nbe used to discriminate face images. In this paper we present this general\nframework, first of its kind for facial verification, and tailor it to Gaussian\ndistributions. This choice enables the use of a simple linear decision boundary\nthat can be tuned to achieve the desired trade-off between false alarm and\ngenuine acceptance rate, and leads to a loss function that can be written in\nclosed form. Extensive analysis and experimentation on publicly available\ndatasets such as Labeled Faces in the wild (LFW), Youtube faces (YTF),\nCelebrities in Frontal-Profile in the Wild (CFP), and challenging datasets like\ncross-age LFW (CALFW), cross-pose LFW (CPLFW), In-the-wild Age Dataset (AgeDB)\nshow a significant performance improvement and confirms the effectiveness and\nsuperiority of BioMetricNet over existing state-of-the-art methods.\n