Person Re-Identification of Cross-Domain Based on Costume Characteristics

Cross-domain person re-identification (ReID) task has always been a research hotspot. The domain transfer technique based on deep learning alleviates the problem caused by the gap among data distributions. However, the generalization of such kinds of person ReID models are not robust. This paper proposes a novel ReID method to optimize the robustness from the perspective of persons' costume. It can extract hierarchical knowledge features from costume information and apply the feature model to the task of ReID. Specifically, the method first constructs hierarchical labels according to the attribute of costume data. It builds the knowledge model by using the multi-task learning, and performs similarity measurement on multiple ReID datasets. The experimental results show that the proposed method outperforms the latest ReID methods. Therefore, it provides a novel technique for practical applications.