Potential source-information dominated learning for composed cross-modal person re-identification

Min Yang, Xiangyun Zhang, Guannan Dong, Kunyu Wu, Yong Cheng, Aichun Zhu

Abstract: Recent years Person Re-identification(ReID) has attracted attention in industry and acade-mia. Previous works ignore the impact of appearance changes, such as clothing changes, leading to inferior performance in information retrieval under textual descriptions. Hence, we present a novel task called Composed Cross-modal Person Re-identification (CCM-ReID). Notably, there is currently no available dataset containing descriptions of appearance changes. Therefore, we construct a new dataset named CCMReid which is collected from published datasets. Meanwhile, in this paper, we propose a Potential Source-Information Dominated Learning (PSDL) model to excavate underlying essential information and clothing-changing representation by color-grey space contrastive training on source images for robust person ReID. Specifically, we introduce a source-information extraction module to tackle different color gamuts to explore real and abundant source information. To cope with the appearance variations in the open world, we also design potential information-dominated contrastive learning, which leverages different gamut information to greatly explore the latent changes expressed by related captions. The experimental results indicate that our PSDL has achieved excellent performance, and CCMReid dataset will be publicly available to support future research.