Parsing-guided Prompt Design for Clothes-Changing Person Re-Identification with CLIP

Abstract

Clothes-changing person re-identification aims to identify the same person under clothing changes, where conventional re-identification methods suffer from significant performance degradation due to the variation of clothing, which is one of the primary discriminative cues. Recently, CLIP-based approaches leveraging the image–text matching capability of vision-language models (VLMs) have been applied to person re-identification. However, their application to clothes-changing scenarios remains limited, and their reliance on global semantic representations makes it difficult to effectively learn local body-part features that are crucial for this task. To address this issue, this paper proposes a Parsing-guided Prompt Designframework that utilizes human parsing information to effectively learn local body-part features. The proposed method suppresses background noise while preserving only human body region information, thereby encouraging the model to focus on intrinsic human morphological characteristics. In addition, body-part-level visual features are cross-modally aligned with text prompts corresponding to each body part, enabling robust identity discrimination that is less dependent on clothing appearance. Experiments on the representative clothes-changing benchmark PRCCdemonstrate that the proposed method achieves superior performance compared with recent state-of-the-art methods.