Multi-granularity Dynamic Identity Aggregation for Text-based Person Re-identification

Text-based person re-identification(ReID) is a text-to-image retrieval task that searches images of people across non-overlapping cameras with textual descriptions. Due to significant intra-class variations and cross-modality disparities, along with substantial sample noise, it becomes challenging to learn discriminative features. Existing text-based ReID methods tend to focus on learning global representations, which possess limited discriminability and weak robustness to noisy samples. In this paper, we propose a novel Multi-granularity Dynamic Identity Aggregation(MuDIA) learning approach that simultaneously explores intra-modal fine-grained and cross-modal coarse-grained contextual cues for text-based ReID. MuDIA consists of two primary components: Weighted Parts Aggregation for Global Feature (WPAG) and Multi-Granularity Refinement of Local Feature (MGRL). WPAG aims to learn discriminative features by aggregating parts and simultaneously exploring the contextual relations among body parts within each modality. MGRL aligns the features of the two modalities at multi-granularity, eliminating the modality gap at cascade-style levels and gradually reducing the feature gap between the visual domain and the textual domain. Extensive experiments are carried out to evaluate the proposed framework on two text-based ReID datasets, i.e., CUHK-PEDES and ICFG-PEDES.