LLM-based Person Re-Identification Based on Multi-Modal Feature Fusion

Person re-identification continues to face significant challenges in complex real-world scenarios characterized by occlusions, lighting variations, and low image quality, where conventional visual-only approaches often fail to distinguish similar-looking individuals due to their inherent single-modal limitations. To overcome these constraints, we present a novel multi-modal solution that effectively combines visual features with semantically rich textual descriptions generated through BLIP, establishing the first comprehensive text-visual fusion framework for Reid applications. Our approach introduces three key advancements: an attention-based dynamic fusion mechanism that automatically balances modality contributions, a crossmodal contrastive loss that enhances feature alignment, and the integration of large language models (BLIP/BERT) for robust text processing. Experimental validation on CUHK03 demonstrates consistent performance improvements over visual baselines, while our newly created text-annotated dataset extension provides valuable resources for future multi-modal Reid research. These contributions collectively advance the field by bridging the gap between visual and textual modalities while establishing reproducible benchmarks for subsequent studies.