Face Recognition Using Modified OpenFace

Abstract

Face recognition is a widely utilized biometric method due to its natural and non-intrusive approach. Recently, deep learning networks using Triplet Loss have become a common framework for person identification and verification. In this paper, we present a new method on how to select appropriate hard-negatives for training using Triplet Loss. We show that, by incorporating pairs which would otherwise have been discarded yields better accuracy and performance. We also applied Adaptive Moment Estimation algorithm to mitigate the risk of early convergence due to the additional hard-negative pairs. In LFW verification benchmark, we managed to achieve an accuracy of 0.955 and AUC of 0.989 as opposed to 0.929 and 0.973 in the original OpenFace.