Scholarlink

In summary, our innovative retrieval system for interactive video search, developed for the VBS 2025 competition, significantly elevates the user experience through the utilization of LLMs. By integrating LLMs for advanced query expansion techniques, we effectively address ambiguities and broaden search parameters, resulting in enhanced retrieval accuracy. Our proposed multimodal search framework is designed to support a variety of input types, including text queries, visual data, object filtering, and visual queries generated with Stable Diffusion, providing users with a flexible and intuitive search experience. Furthermore, our dynamic temporal search strategy offers a comprehensive evaluation of frame relevance, surpassing traditional methods and delivering a richer and more effective video retrieval experience for users.

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

NII-UIT at VBS2025: Multimodal Video Retrieval with LLM Integration and Dynamic Temporal Search

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Bao Tran Gia,
  • Tuong Bui Cong Khanh,
  • Tam Le Thi Thanh,
  • Thuyen Tran Doan,
  • Khiem Le,
  • Tien Do,
  • Tien-Dung Mai,
  • Thanh Duc Ngo,
  • Duy-Dinh Le,
  • Shin’ichi Satoh

摘要

In summary, our innovative retrieval system for interactive video search, developed for the VBS 2025 competition, significantly elevates the user experience through the utilization of LLMs. By integrating LLMs for advanced query expansion techniques, we effectively address ambiguities and broaden search parameters, resulting in enhanced retrieval accuracy. Our proposed multimodal search framework is designed to support a variety of input types, including text queries, visual data, object filtering, and visual queries generated with Stable Diffusion, providing users with a flexible and intuitive search experience. Furthermore, our dynamic temporal search strategy offers a comprehensive evaluation of frame relevance, surpassing traditional methods and delivering a richer and more effective video retrieval experience for users.