Angular Discriminative Deep Feature Learning for Face Verification

Abstract

Thanks to the development of deep Convolutional Neural Network (CNN), face verification has achieved great success rapidly. Specifically, Deep Distance Metric Learning (DDML), as an emerging area, has achieved great improvements in computer vision community. Softmax loss is widely used to supervise the training of most available CNN models. Whereas, feature normalization is often used to compute the pair similarities when testing. In order to bridge the gap between training and testing, we require that the intra-class cosine similarity of the inner-product layer before softmax loss is larger than a margin in the training step, accompanied by the supervision signal of softmax loss. To enhance the discriminative power of the deeply learned features, we extend the intra-class constraint to force the intra-class cosine similarity larger than the mean of nearest neighboring inter-class ones with a margin in the normalized exponential feature projection space. Extensive experiments on Labeled Face in the Wild (LFW) and Youtube Faces (YTF) datasets demonstrate that the proposed approaches achieve competitive performance for the open-set face verification task.