A realistic approach to generate masked faces applied on two novel masked face recognition data sets

The COVID-19 pandemic raises the problem of adapting face recognition systems\nto the new reality, where people may wear surgical masks to cover their noses\nand mouths. Traditional data sets (e.g., CelebA, CASIA-WebFace) used for\ntraining these systems were released before the pandemic, so they now seem\nunsuited due to the lack of examples of people wearing masks. We propose a\nmethod for enhancing data sets containing faces without masks by creating\nsynthetic masks and overlaying them on faces in the original images. Our method\nrelies on SparkAR Studio, a developer program made by Facebook that is used to\ncreate Instagram face filters. In our approach, we use 9 masks of different\ncolors, shapes and fabrics. We employ our method to generate a number of\n445,446 (90%) samples of masks for the CASIA-WebFace data set and 196,254\n(96.8%) masks for the CelebA data set, releasing the mask images at\nhttps://github.com/securifai/masked_faces. We show that our method produces\nsignificantly more realistic training examples of masks overlaid on faces by\nasking volunteers to qualitatively compare it to other methods or data sets\ndesigned for the same task. We also demonstrate the usefulness of our method by\nevaluating state-of-the-art face recognition systems (FaceNet, VGG-face,\nArcFace) trained on our enhanced data sets and showing that they outperform\nequivalent systems trained on original data sets (containing faces without\nmasks) or competing data sets (containing masks generated by related methods),\nwhen the test benchmarks contain masked faces.\n