A Separating Embedding Space Based Relation Network with RGB Modality Only for Cloth-Changing Person Re-Identification

Jian Xu, Bo Liu, Yanshan Xiao

Abstract: The Cloth-changing person re-identification (CC-ReID) is more challenging than person re-identification (Re-ID) since the cloth-relevant features are unreliable. The current CC-ReID methods usually utilize some human parsing techniques such as semantic segmentation to guide the model to learn more cloth-irrelevant feature cues. However, the human parsing models are not necessarily reliable. For this, we propose a Separating Embedding Space based Relation Network (SESRN) for CC-ReID. Firstly, we use pairs of images as input of the model and consider the relation between them in SESRN instead of a single image that the existing CC-ReID models use. Secondly, we propose to separate the common feature embedding space outputted from the common backbone into two embedding subspaces including the cloth-irrelevant feature embedding subspace and cloth-related feature embedding subspace without using any human parsing techniques since we think that they are unreliable and introduce noise into model. Thirdly, we generate the different feature map weights on different subspaces or different comparison pairs on the same subspace, which is a simple and effective feature map visualization and analysis framework in CC-ReID. Finally, the extensive experiments show the effectiveness and robustness of our method.