A Survey on Text-Based Person Search

Text-Based Person Search (TBPS) is a fundamental problem in intelligent surveillance and multimedia retrieval, aiming to identify a target pedestrian in large-scale image galleries using free-form natural language descriptions. Despite rapid advances, existing research is scattered across diverse models, datasets, and evaluation protocols, underscoring the need for a unified and critical survey. This work provides a systematic review of TBPS by formalizing the problem setting and presenting a structured taxonomy of representative methods along four key technical dimensions: (1) external knowledge-based approaches that leverage semantic priors to alleviate the inherent modality gap between visual and textual data; (2) learning objectives that employ diverse and evolving loss functions to guide cross-modal representation learning; (3) advanced encoder architectures for robust unimodal feature extraction; and (4) modality interaction mechanisms enabling fine-grained visual-textual alignment. We further summarize benchmark datasets, evaluation protocols, and comparative performance, and review related tasks that extend the TBPS paradigm. Finally, we outline open challenges and future research directions.