Multi-Level Semantic-Guided Framework for Cloth-Changing Person Re-Identification

Abstract

Cloth-changing person re-identification (CC-ReID) aims to match persons who change clothing across multiple surveillance cameras. Recent approaches strive to extract clothing-agnostic features by utilizing biological information, including skeleton, texture, gait, and 3D data. However, most methods rely on additional features at a single aspect, leading to a lack of comprehensive understanding of concepts and semantics. This limitation introduces biases and restricts both accuracy and functionality, thereby diminishing their effectiveness in handling variations in clothing. To alleviate this problem, we propose a Multi-level Semantic-Guided (MSG) framework that integrates contextual and fine-grained visual information to eliminate clothing variance across both conceptual and pixel dimensions. This innovative solution consists of two key components: the Contextual Semantic Guidance (CSG) module adeptly utilizes textual features from clothing descriptions to decouple clothing concepts at a higher semantic level. In contrast, the Low-Level Robust Feature Disentanglement (LRD) module meticulously analyzes images featuring clothing to disentangle texture information at the granular pixel level, and the integration of a momentum update mechanism significantly bolsters the model’s robustness. The two approaches collaboratively eliminate clothing information by interacting across different semantic levels. Extensive experiments demonstrate the effectiveness of our method, achieving new state-of-the-art performance on several popular CC-ReID benchmarks. Our code will be available on GitHub at https://github.com/ShijuanHuang/MSG.