Deep learning, in particular Convolutional Neural Network (CNN), has achievedpromising results in face recognition recently. However, it remains an openquestion: why CNNs work well and how to design a 'good' architecture. Theexisting works tend to focus on reporting CNN architectures that work well forface recognition rather than investigate the reason. In this work, we conductan extensive evaluation of CNN-based face recognition systems (CNN-FRS) on acommon ground to make our work easily reproducible. Specifically, we use publicdatabase LFW (Labeled Faces in the Wild) to train CNNs, unlike most existingCNNs trained on private databases. We propose three CNN architectures which arethe first reported architectures trained using LFW data. This paperquantitatively compares the architectures of CNNs and evaluate the effect ofdifferent implementation choices. We identify several useful properties ofCNN-FRS. For instance, the dimensionality of the learned features can besignificantly reduced without adverse effect on face recognition accuracy. Inaddition, traditional metric learning method exploiting CNN-learned features isevaluated. Experiments show two crucial factors to good CNN-FRS performance arethe fusion of multiple CNNs and metric learning. To make our work reproducible,source code and models will be made publicly available.
展开▼