Deep Regression Neural Network for End-to-End Person Re-Identification

Abstract

Person re-identification can be seen as a process of open set recognition. Usually, the deep learning models consider the person re-identification model as a classification model with a softmax layer. However, the softmax layer cannot be extended to unknown classes because of its closed nature, so the classification model is just regarded as the feature extractor. To overcome the problem mentioned above and make the person re-identification process end-to-end, this paper cast the person re-identification into a regression process and calculates the probability that persons in the images belong to the same identity. First, this paper proposes a deep regression model, named deep regression neural network integrating adaptive multi-attribute fusion method (DRNN-AMAF), which can make the person re-identification as regression analysis. Second, attributes are taken as the basis of this model for calculating the probability of persons belonging to the same identity, and each attribute corresponds to each branch of the deep regression neural network. Finally, hard labels of multiple attributes are adaptively fused into a soft label by the proposed multi-label fusion method based on the idea of Bayesian inference, which makes the attribute labels suitable for regression tasks. The comprehensive experiments on available public databases are conducted, and the experimental results show that our model produces competitive performance compared with the state-of-the-art approaches.