Scholarlink

Text-based pedestrian search (TBPS) in full images aims to locate a target person in uncropped images using natural language queries. However, in complex scenes with multiple pedestrians often lead to ambiguous detections and misalignment between language and vision. To address these challenges, we propose UPD-TBPS, a novel framework composed of three modules: Multi-granularity Uncertainty Estimation (MUE), Prototype-based Uncertainty Decoupling (PUD), and Cross-modal Re-identification (ReID). MUE reduces early-stage detection ambiguity via confidence-scored multi-view queries. PUD decouples multi-level visual semantics and mines coarse-to-fine prototypes to guide grounding. ReID leverages uncertainty-aware features to improve final retrieval. Experiments on two TBPS benchmarks adapted to full-image settings demonstrate the effectiveness of our framework.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Uncertainty-Aware Prototype Semantic Decoupling for Text-Based Person Search in Full Images

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Zengli Luo,
  • Canlong Zhang,
  • Zhixin Li,
  • Zhiwen Wang,
  • Chunrong Wei

摘要

Text-based pedestrian search (TBPS) in full images aims to locate a target person in uncropped images using natural language queries. However, in complex scenes with multiple pedestrians often lead to ambiguous detections and misalignment between language and vision. To address these challenges, we propose UPD-TBPS, a novel framework composed of three modules: Multi-granularity Uncertainty Estimation (MUE), Prototype-based Uncertainty Decoupling (PUD), and Cross-modal Re-identification (ReID). MUE reduces early-stage detection ambiguity via confidence-scored multi-view queries. PUD decouples multi-level visual semantics and mines coarse-to-fine prototypes to guide grounding. ReID leverages uncertainty-aware features to improve final retrieval. Experiments on two TBPS benchmarks adapted to full-image settings demonstrate the effectiveness of our framework.