Scholarlink

The shift towards a 3D-model-based workflow in manufacturing within the context of Industry 4.0 is challenged by 2D engineering drawings (EDs) still containing critical product manufacturing information (PMI) as the sole source of truth. Manually transferring PMI data from EDs to 3D-CAD-models is a time-consuming, error-prone process that requires expert knowledge. This paper presents a pipeline that automates the extraction of 2D drafting annotations. It utilizes location data relative to the projected 3D-CAD-model to reference the associated geometry, enhancing the 3D model with converted 3D PMI data. The approach leverages deep learning-based object detection (finetuned YOLOv8-OBB) for PMI extraction, a novel Optical Character Recognition (OCR) model (finetuned GOT-OCR) for text retrieval, and a heuristic algorithm for mapping 2D drafting annotations to respective 3D geometry. The proposed method is evaluated on both synthetic and industry EDs, achieving high accuracy in 2D drafting annotation detection (91.3% mAP), text recognition (95.1%), and PMI placement accuracy (89.5%). The proposed method significantly reduces the manual effort required for PMI transfer, facilitating seamless digital workflows in manufacturing.

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

Object Detection in Engineering Drawings: An End-to-End Conversion Pipeline from 2D Drafting Annotations to 3D Product Manufacturing Information

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Kilian Eisert,
  • Tim Benjamin Hoffmann,
  • Oliver Lohse

摘要

The shift towards a 3D-model-based workflow in manufacturing within the context of Industry 4.0 is challenged by 2D engineering drawings (EDs) still containing critical product manufacturing information (PMI) as the sole source of truth. Manually transferring PMI data from EDs to 3D-CAD-models is a time-consuming, error-prone process that requires expert knowledge. This paper presents a pipeline that automates the extraction of 2D drafting annotations. It utilizes location data relative to the projected 3D-CAD-model to reference the associated geometry, enhancing the 3D model with converted 3D PMI data. The approach leverages deep learning-based object detection (finetuned YOLOv8-OBB) for PMI extraction, a novel Optical Character Recognition (OCR) model (finetuned GOT-OCR) for text retrieval, and a heuristic algorithm for mapping 2D drafting annotations to respective 3D geometry. The proposed method is evaluated on both synthetic and industry EDs, achieving high accuracy in 2D drafting annotation detection (91.3% mAP), text recognition (95.1%), and PMI placement accuracy (89.5%). The proposed method significantly reduces the manual effort required for PMI transfer, facilitating seamless digital workflows in manufacturing.