Feature Norm Regularization and Class Center Norm Standardization for Person Re-Identification

Abstract

Person re-identification (ReID) and face recognition both address the problem of calculating the similarity of two images, but the cross-entropy loss functions they use have significant differences in form. By analyzing the differences we believe it is necessary to introduce consistency requirements for feature norms and class-center norms in ReID domain to enhance the generalization performance of the algorithm. To address the issue of the changing mean value of feature norms during training, we establish a feature norm observer, which allows us to utilize the estimated real-time mean value to form a feature norm regularization loss for the current training batch. We employ a similar method to observe the gradient norm of the cross-entropy loss and that of the feature norm regularization loss. By dynamically adjusting the loss weight, we ensure that the gradient intensities from two losses are aligned. Regarding the class center norm, after analyzing the necessity of its consistency, we propose a simple method that uniformly resets all class center norms to their average value at the end of the current training batch. The experiments demonstrate the effectiveness of the proposed method. After applying the proposed method for OSNet, the increment of mAP on Market1501, CUHK03 and MSMT17 ranges from 0.8% to 4.5%, and that of rank-1 ranges from 0.2% to 3.2%.