Scholarlink

This study explores the use of few-shot learning for landmark recognition by integrating a modified ResNet-18 embedding network with a prototypical network. The model was trained and evaluated on the Google Landmarks Dataset, making use of episodic training to simulate real-world few-shot learning scenarios. The training process achieved a high accuracy of 97.67%, while testing on previously unseen classes resulted in an accuracy of 92.00%, demonstrating strong generalization capabilities. The findings highlight the usefulness of the proposed approach in recognizing landmarks with limited labeled data. Future work will focus on improving dataset balance, exploring advanced meta-learning architectures, and incorporating domain-specific knowledge to further enhance performance.

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

Few-Shot Learning for Landmarks Classification

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Aarabhi Anand,
  • B. Ananya,
  • Anusha Kallur,
  • Bhavana Bandi,
  • K. Gautham,
  • Bhaskarjyoti Das

摘要

This study explores the use of few-shot learning for landmark recognition by integrating a modified ResNet-18 embedding network with a prototypical network. The model was trained and evaluated on the Google Landmarks Dataset, making use of episodic training to simulate real-world few-shot learning scenarios. The training process achieved a high accuracy of 97.67%, while testing on previously unseen classes resulted in an accuracy of 92.00%, demonstrating strong generalization capabilities. The findings highlight the usefulness of the proposed approach in recognizing landmarks with limited labeled data. Future work will focus on improving dataset balance, exploring advanced meta-learning architectures, and incorporating domain-specific knowledge to further enhance performance.