Scholarlink

<p>Text-driven person retrieval aims to find a particular person within an array of images using a natural language description of that individual. The main challenge in this task lies in calculating a similarity score that measures the likeness between the person’s image and the given description. This procedure involves uncovering the intricate, underlying connections between various levels of image sub-regions and textual phrases. Previous efforts have sought to tackle this issue by utilizing individual pre-trained models for visual and textual data to extract features. However, these methods lack essential alignment capabilities regarding efficient multimodal data matching. Moreover, these approaches rely on prior information to investigate explicit part alignments, potentially causing the disruption of information within the same modality. To increase the performance of existing state-of-the-art text-to-image retrieval systems, we proposed a novel technique aimed at enhancing the effectiveness of the text-image pedestrian matching pipeline, referred to as Complete Feature Aggregation and Cross-Knowledge Conversion (CFA-CKC), which integrates diverse supervisory signals from various textual and visual perspectives. Our proposed algorithm achieves new state-of-the-art results on all three public datasets, including CUHK-PEDES, ICFG-PEDES, and RSTPReid. Experimental results show that our CFA-CKC outperforms the state-of-the-art methods by up to 6.0% in rank-1 accuracy. The code and pre-trained models are available at this link <a href="https://github.com/AIVIETNAMResearch/Pedestrian-Matching">https://github.com/AIVIETNAMResearch/Pedestrian-Matching</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

An efficient framework for text-to-image retrieval using complete feature aggregation and cross-knowledge conversion

  • 期刊论文

  • 发表日期：

    2025年9月26日

查看全文
  • Bach Hoang Ngo,
  • Minh-Hung An,
  • Khoa Nguyen Tho Anh,
  • Minh-Duc Bui,
  • Quang-Vinh Dinh,
  • Vinh Dinh Nguyen

摘要

Text-driven person retrieval aims to find a particular person within an array of images using a natural language description of that individual. The main challenge in this task lies in calculating a similarity score that measures the likeness between the person’s image and the given description. This procedure involves uncovering the intricate, underlying connections between various levels of image sub-regions and textual phrases. Previous efforts have sought to tackle this issue by utilizing individual pre-trained models for visual and textual data to extract features. However, these methods lack essential alignment capabilities regarding efficient multimodal data matching. Moreover, these approaches rely on prior information to investigate explicit part alignments, potentially causing the disruption of information within the same modality. To increase the performance of existing state-of-the-art text-to-image retrieval systems, we proposed a novel technique aimed at enhancing the effectiveness of the text-image pedestrian matching pipeline, referred to as Complete Feature Aggregation and Cross-Knowledge Conversion (CFA-CKC), which integrates diverse supervisory signals from various textual and visual perspectives. Our proposed algorithm achieves new state-of-the-art results on all three public datasets, including CUHK-PEDES, ICFG-PEDES, and RSTPReid. Experimental results show that our CFA-CKC outperforms the state-of-the-art methods by up to 6.0% in rank-1 accuracy. The code and pre-trained models are available at this link https://github.com/AIVIETNAMResearch/Pedestrian-Matching.