基于身份与属性感知对齐的文本行人重识别

Text-Based Person Re-Identification via Identity and Attribute-Aware Alignment

摘要: 文本行人重识别(Text-based Person Re-Identification, TextReID)旨在根据自然语言描述在大规模行人图像库中检索对应个体，在智能安防等场景中具有重要应用价值。近年来，大规模的图文预训练模型(如：Contrastive Language-Image Pre-Training, CLIP)被广泛迁移到该任务中，然而CLIP默认采用的实例级对比学习目标与行人重识别数据的真实分布存在结构性错配，同时文本侧对行人属性语义建模不足，导致检索性能不稳定、泛化能力受限。针对上述问题，本文提出一种结构感知CLIP适配框架，从目标空间与语义空间两个层面对CLIP进行结构性修正。具体地，提出身份感知对齐损失(Identity-Aware Alignment, IAA)，将实例级图文对齐提升为身份级分布对齐，避免同一身份样本被误判为负样本；同时提出属性感知掩码建模(Attribute-Aware Masked Modeling, AAMM)，在文本编码阶段重点建模与行人属性相关的词汇语义，增强判别性文本表征。在多个公开数据集上的实验结果表明，我们的方法在Rank-1、mAP等指标上均显著优于现有方法，验证了所提框架的有效性与泛化能力。

Abstract: Text-based Person Re-Identification (TextReID) aims to retrieve corresponding pedestrian images from large-scale galleries based on natural language descriptions and plays an important role in intelligent surveillance. Recently, CLIP has been widely adopted for this task; however, its instance-level contrastive objective mismatches the true data distribution where multiple images and texts correspond to the same identity. Meanwhile, insufficient modeling of attribute-related semantics in textual descriptions further limits its discriminative capability. To address these issues, we propose a structure-aware CLIP adaptation framework that corrects the structural mismatches from both objective and semantic perspectives. Specifically, we introduce an Identity-Aware Alignment (IAA) loss to align identity-level distributions instead of individual instances, preventing same-identity samples from being mistakenly treated as negatives. Moreover, an Attribute-Aware Masked Modeling (AAMM) module is designed to emphasize attribute-related tokens during text encoding, thereby enhancing discriminative textual representations. Extensive experiments on multiple public benchmarks demonstrate that the proposed method significantly outperforms existing approaches in terms of Rank-1 and mAP, validating its effectiveness and generalization ability.