Scholarlink

<p>Widely regarded as a fundamental challenge in computer vision, local image feature matching has been thoroughly studied in recent years. However, semi-dense methods often suffer from an insufficient number of matches during coarse matching, while dense methods, though more thorough, demand significant computational resources and exhibit low efficiency. To compensate for these deficiencies, we propose Increased Matching Amount Semi-Dense Matching (IMAmatch), a novel method that selectively increases the number of matches in key image regions to significantly enhance both accuracy. IMAmatch improves feature extraction through a multi-scale focused linear attention mechanism and boosts matching accuracy by densifying correspondences in key areas. The resulting matches are further refined using a consistency check between regression and classification outputs. Finally, the method is optimized with a differentiable relative pose estimation strategy based on bundle adjustment. Extensive experiments across multiple datasets demonstrate that IMAmatch achieves performance on par with dense matching methods, while being notably more memory-efficient and faster. To support reproducibility, the code will be available at: <a href="https://github.com/LiaoYun0x0/IMAmatch">https://github.com/LiaoYun0x0/IMAmatch</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

Semi-dense feature matching with increased matching amount

  • 期刊论文

  • 发表日期：

    2025年8月1日

查看全文
  • Yide Di,
  • Yun Liao,
  • Yunan Liu,
  • Mingyu Lu,
  • Qing Duan,
  • Junhui Liu

摘要

Widely regarded as a fundamental challenge in computer vision, local image feature matching has been thoroughly studied in recent years. However, semi-dense methods often suffer from an insufficient number of matches during coarse matching, while dense methods, though more thorough, demand significant computational resources and exhibit low efficiency. To compensate for these deficiencies, we propose Increased Matching Amount Semi-Dense Matching (IMAmatch), a novel method that selectively increases the number of matches in key image regions to significantly enhance both accuracy. IMAmatch improves feature extraction through a multi-scale focused linear attention mechanism and boosts matching accuracy by densifying correspondences in key areas. The resulting matches are further refined using a consistency check between regression and classification outputs. Finally, the method is optimized with a differentiable relative pose estimation strategy based on bundle adjustment. Extensive experiments across multiple datasets demonstrate that IMAmatch achieves performance on par with dense matching methods, while being notably more memory-efficient and faster. To support reproducibility, the code will be available at: https://github.com/LiaoYun0x0/IMAmatch.