Scholarlink

The widespread sharing of visual data online has created challenges in querying and extracting relevant information efficiently. This led to the establishment of the AI Challenge 2024, which aims to develop systems capable of processing queries and accurately returning event images from a large dataset of multi-genre videos. Our team addressed this by creating an interactive system leveraging models such as CLIP, GPT-4o, PaddleOCR, and Whisper to generate precise embeddings and enable efficient data retrieval from diverse visual formats. The system is supported by an GUI with features such as semantic search, OCR-based and voice-based queries, AI-generated image queries, query enhancement, and advanced video preprocessing. Through this approach, our system reduced the unimportant data by 21% and achieved an accuracy of 81.54% for the correct answers among the top 10 responses ( https://github.com/NghiaZun/AIthena_Statistic ). This result showcases the system’s potential for improving query response accuracy and efficiency in large-scale visual data preprocessing.

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 Interactive System For Visual Data Retrieval From Multimodal Input

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Tu Van Nguyen,
  • Nghia Trung Duong,
  • Nhan Thanh Pham,
  • Thanh Xuan Luong,
  • Dang Duy Bui

摘要

The widespread sharing of visual data online has created challenges in querying and extracting relevant information efficiently. This led to the establishment of the AI Challenge 2024, which aims to develop systems capable of processing queries and accurately returning event images from a large dataset of multi-genre videos. Our team addressed this by creating an interactive system leveraging models such as CLIP, GPT-4o, PaddleOCR, and Whisper to generate precise embeddings and enable efficient data retrieval from diverse visual formats. The system is supported by an GUI with features such as semantic search, OCR-based and voice-based queries, AI-generated image queries, query enhancement, and advanced video preprocessing. Through this approach, our system reduced the unimportant data by 21% and achieved an accuracy of 81.54% for the correct answers among the top 10 responses ( https://github.com/NghiaZun/AIthena_Statistic ). This result showcases the system’s potential for improving query response accuracy and efficiency in large-scale visual data preprocessing.