Contrapositive Margin Softmax Loss for Face Verification

Abstract

The performance of face recognition has been boosted by the features extracted from deep convolutional neural networks. Ideal features should have minimum intra-class variations and maximum inter-class variations. The most commonly used loss function for classification, softmax loss, however, does not necessarily learn features discriminative enough. Large margin classifiers have nice generalization properties in statistical machine learning. These properties have lead to the application of margin to deep learning in recent years. We hereby propose a new loss function called Contrapositive Margin Softmax loss for face verification task, which helps to learn invariant and discriminative features by introducing margins to both target logits and maximum negative logits of softmax loss. Competitive results on LFW (99.28%) and YTF (95.34%) demonstrate the effectiveness of our approach.