Scholarlink

Searching for specific objects within video footage is a challenging task that has attracted the attention of many researchers. Historically, traditional vision algorithms have focused on improving the accuracy of locating specific objects or contexts in individual images or image sequences. Despite these advancements, searching for specific contexts in video recordings using natural language remains a challenging task. With the increasing utilization of Large Language Models (LLMs) across various domains, this paper introduces a novel architecture designed to query specific objects in video recordings based on their characteristics. This paper evaluates the performance of GPT-4, Llama 2, and Zephyr in generating relevant queries for NoSQL databases. This investigation aims to determine the most effective approach between Zero-shot (no instructions) and Few-shot (example prompts and example results), for generating optimal results. Our findings demonstrate a significant improvement in generating NoSQL queries when applying the Few-shot paradigm compared with the Zero-shot paradigm. In addition, the experiments demonstrate that a small open-source model (Zephyr 7b) can outperform a 70 billion parameters LLM (Llama 2) in a Few-shot setting.

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

Evaluating LLMs for Effective Object Search in Video Data

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Rui David,
  • Jacky Lee,
  • Darren Constantine,
  • Md Fardin Rahman,
  • Qusay H. Mahmoud

摘要

Searching for specific objects within video footage is a challenging task that has attracted the attention of many researchers. Historically, traditional vision algorithms have focused on improving the accuracy of locating specific objects or contexts in individual images or image sequences. Despite these advancements, searching for specific contexts in video recordings using natural language remains a challenging task. With the increasing utilization of Large Language Models (LLMs) across various domains, this paper introduces a novel architecture designed to query specific objects in video recordings based on their characteristics. This paper evaluates the performance of GPT-4, Llama 2, and Zephyr in generating relevant queries for NoSQL databases. This investigation aims to determine the most effective approach between Zero-shot (no instructions) and Few-shot (example prompts and example results), for generating optimal results. Our findings demonstrate a significant improvement in generating NoSQL queries when applying the Few-shot paradigm compared with the Zero-shot paradigm. In addition, the experiments demonstrate that a small open-source model (Zephyr 7b) can outperform a 70 billion parameters LLM (Llama 2) in a Few-shot setting.