Evidence-guided Feature Learning for Text-to-Image Person Re-identification

Text-to-Image Person Re-identification aims to retrieve person images based on natural language descriptions. Existing methods primarily rely on deep feature mapping to learn cross-modal consistency but often overlook the crucial visual-semantic evidence associations between text and images, leading to less interpretable matching processes. To address this issue, we propose an Evidence-guided Feature Learning framework to enhance the accuracy and robustness of cross-modal alignment. Specifically, Furthermore, we introduce an Evidence-guided Feature Aggregation strategy to enhance representation capability by fusing salient local features. Additionally, we propose an Evidence Consistency Loss to constrain the stability and generalization ability of text-to-image matching. Experimental results on CUHK-PEDES, ICFG-PEDES demonstrate that the proposed method significantly outperforms existing state-of-the-art approaches in retrieval accuracy.