Scholarlink

<p>Traditional symbol spotting relies on CAD drawings with precise vector data, which require expensive expert annotation. Spotting symbols directly from CAD images reduces these costs but loses structural information, necessitating highly precise pixel-wise keypoint localization. However, standard heatmap methods suffer from Maximum Value Drift (MVD) and quantization errors induced by downsampling. This study addresses these challenges by proposing a pixel-wise keypoint location technique based on Progressive Gaussian Kernels (PGK), which balances training efficiency and location accuracy. To mitigate quantization errors, a local offset is introduced into the heatmap-based point localization method. Finally, an error-correcting algorithm robustly assembles the detected keypoints into complete rectangular symbols. Evaluated on our newly released CAD image dataset, our method significantly outperforms state-of-the-art baselines, achieving an F1 score of 0.84 and reducing the Averaged Pixel Error for Keypoints (APEK) to 4.36.</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

Accurate pixel-wise keypoint localization for rectangle symbol spotting in CAD images

  • 期刊论文

  • 发表日期：

    2026年7月27日

查看全文
  • Jiaxin Deng,
  • Zhen Huang,
  • Junbiao Pang,
  • Zailin Dong,
  • Mengyuan Zhu

摘要

Traditional symbol spotting relies on CAD drawings with precise vector data, which require expensive expert annotation. Spotting symbols directly from CAD images reduces these costs but loses structural information, necessitating highly precise pixel-wise keypoint localization. However, standard heatmap methods suffer from Maximum Value Drift (MVD) and quantization errors induced by downsampling. This study addresses these challenges by proposing a pixel-wise keypoint location technique based on Progressive Gaussian Kernels (PGK), which balances training efficiency and location accuracy. To mitigate quantization errors, a local offset is introduced into the heatmap-based point localization method. Finally, an error-correcting algorithm robustly assembles the detected keypoints into complete rectangular symbols. Evaluated on our newly released CAD image dataset, our method significantly outperforms state-of-the-art baselines, achieving an F1 score of 0.84 and reducing the Averaged Pixel Error for Keypoints (APEK) to 4.36.