Multi-feature balanced network for clothes-changing person re-identification

Clothes-changing person re-identification (CC-ReID) is pivotal in long-term scenes, especially when involving significant variations in clothes. The principal challenge of this task lies in the extraction of clothes-irrelevant features. Currently, most methods alleviate the interference caused by clothing changes through separating and identifying body regions. However, these approaches could not fully utilize all the useful information in pedestrian images. In this work, a novel Multi-features Balanced Network (MBNet) is proposed for improving the robustness of the CC-ReID model by exploiting clothing-unrelated features, consisting of a global branch, a clothing-unrelated branch, and a mask branch. Specifically, to highlight clothing-unrelated clues, a knowledge transfer module (KTM) is first designed. Then, the clothing-unrelated branch only receives images that are unaffected or less affected clothing to resist clothes-changing. Besides, a feature attention module (FAM) is introduced in this branch, which can suppress background clutter and extract discriminative fine-grained features. Finally, a cross fusion module (CFM) is used to integrate more contextual information and mine more clothing-independent and pose features in the mask branch. Three branches are combined to perform CC-ReID. Extensive experiments on three popular synthetic and realistic datasets show that the superiority of the proposed approach, achieving a Rank-1/mAP accuracy of 44.6%/22.7%, 58.3%/57.9%, 87.2%/84.0%.