LinArc - Deep Face Recognition Using LinCos And ArcFace

Data is overflowing day by day. The use of face recognition is rapidly picking up pace due to the boom in data and available computation power. The research done in this field is at an unimaginable pace, and accuracies of more than 99% have been achieved, which are possibly less only by Baye’s error. However, there is still room for experimentation. This paper tries to build a model by mixing two novel ideas of face recognition - ArcFace and LinCos. In this paper, the target is to manipulate the Additive Angular Margin Loss used by ArcFace by incorporating the ideas of LinCos. We re-train the pre-trained ArcFace model using Mobile FaceNet with a modified loss function. The results suggest that our model optimizes at a faster rate as compared to the ArcFace and LinCos models.