Multi-Loss Weighting for Person Re-Identification Based on Deep Learning

Abstract

Person re-identification is an important component to realize various image recognition systems, e.g., person tracking system by utilizing multiple cameras. Person re-identification based on deep learning has achieved high performance, and cross-entropy loss or triplet loss is generally used as the loss function. In recent years, a linear summation of both loss functions has been attracting attention as an approach to person re-identification. However, when loss functions with different properties are used at the same time, a method of synthesis by weighted linear summation that takes into account the effect of the loss function on the other loss function is necessary. To overcome above problems, in this paper, a method that automatically adjusts the weights of loss functions during learning is proposed.