Deep Face Recognition with Center Invariant Loss

Abstract

Convolutional Neural Networks (CNNs) have been widely used for face recognition and got extraordinary performance with large number of available face images of different people. However, it is hard to get uniform distributed data for all people. In most face datasets, a large proportion of people have few face images. Only a small number of people appear frequently with more face images. These people with more face images have higher impact on the feature learning than others. The imbalanced distribution leads to the difficulty to train a CNN model for feature representation that is general for each person, instead of mainly for the people with large number of face images. To address this challenge, we proposed a center invariant loss which aligns the center of each person to enforce the learned features to have a general representation for all people. The center invariant loss penalizes the difference between each center of classes. With center invariant loss, we can train a robust CNN that treats each class equally regardless the number of class samples. Extensive experiments demonstrate the effectiveness of the proposed approach. We achieve state-of-the-art results on LFW and YTF datasets.