One-Shot Face Recognition Based on Multiple Classifiers Training

Abstract

One-shot face recognition is a challenging problem which requires recognizing novel identities from only one seen face image. One-shot classes are simply neglected because of the lack of training samples. Therefore, these classes contribute less to the improvement of face recognition performance. The main goal of one-shot face recognition task is to use the novel face samples to enhance the ability of network not only in close-set classify, but also in open-set face verification. In this paper, Base data and Novel data is trained separately with two classifiers to reduce the impact of data imbalance. We propose Confidence Constrain Loss to train classifiers in parallel and get better classifiers fusion in the test phase. Besides, we use data augmentation with 3D face reconstruction to obtain a variety of oneshot set's training samples. Thus, our method can effectively increase the recognition accuracy in the novel set without reducing recognition accuracy in base set. Experiments on MS-celeb-1M low-shot dataset demonstrate that our method achieve state-of-the-art which has 98.90% coverage at precision=99% without using external data.