DNA: Improving text-based person search through distillation learning, negated relation-aware learning, and augmented...

Text-based person search (TPS) retrieves a target person from an image gallery using natural-language queries. While widely studied, TPS still struggles with the fundamental problem of aligning visual and textual modalities in a shared representation space. We present DNA, a TPS framework that leverages distillation-based dual-model learning, where a momentum teacher model supervises a CLIP-based student model. DNA further distinguishes itself by explicitly modeling negative expressions in captions-an important yet underexplored source of relational information. To fully exploit both global and fine-grained cues, DNA incorporates negated relation-aware learning and augmented representation learning, enabling more discriminative feature alignment across modalities. Our design pursues three goals: (i) reducing the visual-textual feature gap; (ii) capturing subtle local semantics overlooked by global features; and (iii) integrating negative descriptors to enhance inference reliability. Experimental results demonstrate that DNA achieves state-of-the-art performance on three benchmark datasets, reaching 76.0% Rank-1 and 67.6% mAP on CUHK-PEDES, 67.7% Rank-1 and 40.3% mAP on ICFG-PEDES, and 65.3% Rank-1 and 51.5% mAP on RSTPReID, consistently outperforming recent CLIP-based TPS methods.