Rethinking Feature Learning Approach for Face Verification

Convolutional neural networks (CNNs) have been broadly used in various computer vision tasks. Almost all of them achieved fascinating results, some even surpassed human performance. In the field of face verification, there are also some attempts of utilizing CNNs for better verification rate. Most of these methods use cross-entropy loss. In order to reinforce the discriminability of the learned features, which is especially important for face verification, we introduce a set of supervision signals in this paper. Specifically, the supervision signals can be separated into two groups: one for redundancy reduction, named decovariance loss, which forces the learned features to diverge; the other called intra-inter center loss(IICL), which achieves both intra-class compactness and interclass discrimination. Experiments validate that the proposed method is both effective and of necessity to the purpose of learning discriminative features. It achieved 98.8% and 93.8% verification accuracy on LFW and YouTube Faces respectively, under the circumstance of using only 0.46M training images.