Enhancing Text-to-Image Person Re-Identification via Discriminative Text Generation

Text-to-image person re-identification aims to retrieve corresponding images based on given text descriptions. The primary challenge lies in achieving fine-grained cross-modal matching between images and texts to distinguish different identities. However, existing text annotations are usually generated individually for single images, causing images of different identities with similar appearances to have similar text descriptions, thereby greatly diminishing the discriminative power of the text. To address this issue, we propose a Discriminative Text Generation method (DTG) based on multimodal large models. Specifically, we perform nearest neighbor search within the original text space to identify cross-identity texts with similar descriptions. We then feed the corresponding image pairs and corresponding texts into a Multimodal Large Language Model (MLLM) to generate more discriminative text descriptions. The experimental results show that when only the generated text descriptions are used, Rank-1 improves by 0.75%. When both the original text descriptions and the generated text descriptions are used, Rank-1 and mAP increase by 1.05% and 0.10%, respectively. The model is able to better identify pedestrians in cross-modal scenarios.