Scholarlink

<p>Person search in unstructured imagery requires jointly performing pedestrian detection and person re-identification (Re-ID), two tasks that inherently compete for computational resources and place conflicting demands on feature alignment. To address this challenge, we propose the All-in-One Person Search (AIOPS) framework, which revisits the traditional OIM-based architecture by unifying and harmonizing feature extraction and representation across heterogeneous backbones. This integration recalibrates inter-module feature propagation and improves detection quality, thereby enhancing the accuracy of subsequent Re-ID. Building on AIOPS, we further develop a Few-Shot Guided Network (FSGN) to better handle challenging and low-data scenarios via meta-learning. FSGN accelerates model adaptation by rapidly extracting discriminative features from only a few examples, while an advanced part-based occlusion attention mechanism and multi-granular feature fusion further strengthen the representational capacity, enabling the recovery of fine details even under severe occlusion and ambiguity. We validate the proposed system through extensive quantitative evaluations on three benchmark datasets: CUHK-SYSU and PRW for person search, and CUB for few-shot learning. The results demonstrate the effectiveness and robustness of our method across diverse settings. The implementation is publicly available at <a href="https://github.com/githubXin89/FSGN">https://github.com/githubXin89/FSGN</a>.</p>

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

All-in-one person search via few-shot guided feature enhancement

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Xiaoqi Xin,
  • Cheng Feng,
  • Dezhi Han

摘要

Person search in unstructured imagery requires jointly performing pedestrian detection and person re-identification (Re-ID), two tasks that inherently compete for computational resources and place conflicting demands on feature alignment. To address this challenge, we propose the All-in-One Person Search (AIOPS) framework, which revisits the traditional OIM-based architecture by unifying and harmonizing feature extraction and representation across heterogeneous backbones. This integration recalibrates inter-module feature propagation and improves detection quality, thereby enhancing the accuracy of subsequent Re-ID. Building on AIOPS, we further develop a Few-Shot Guided Network (FSGN) to better handle challenging and low-data scenarios via meta-learning. FSGN accelerates model adaptation by rapidly extracting discriminative features from only a few examples, while an advanced part-based occlusion attention mechanism and multi-granular feature fusion further strengthen the representational capacity, enabling the recovery of fine details even under severe occlusion and ambiguity. We validate the proposed system through extensive quantitative evaluations on three benchmark datasets: CUHK-SYSU and PRW for person search, and CUB for few-shot learning. The results demonstrate the effectiveness and robustness of our method across diverse settings. The implementation is publicly available at https://github.com/githubXin89/FSGN.