Fine-grained attribute-aware analysis for person re-identification

How and where to learn the discriminative feature has always been a critical issue of person re-identification (re-ID). Most of the previous methods focus on extracting global representation from overall identity image and lack of high-level semantic attribute information. Based on the considerations above, we focus on looking for a way to make people’s attribute play a role in person re-identification task. In this paper, a novel multitask-like network, namely, Attribute-Identity Complementary Network (AICNet), is designed. It contains two branches to learn the identity and attribute feature separately, and a reciprocal interaction process is added to enrich the discrimination of the resulting feature. In addition, in order to better extract high-level semantic attribute information, multi-scale features are combined in the process of attribute extraction. With the help of this attribute-identity complementary strategy, the generated feature can be guided to learn most distinctive attribute feature that is most relevant to identity feature. The present paper carries out extensive experiments on two large-scale datasets, including Market-1501, and DukeMTMC-reID, after which it is found that our attribute-identity complementary framework significantly outperforms the baseline model and achieves competitive performance compared with the state-of-the-art person re-ID methods.