Public Hand‑drawn Circuit Diagram Dataset (GTDB‑HD)

Dataset Overview

• Name: Public Hand‑drawn Circuit Diagram Dataset (GTDB‑HD)
• License: Creative Commons Attribution Share Alike 3.0
• Size: 1K < n < 10K
• Task Categories:
  • Object Detection
  • Image Segmentation
• Languages:
  • English
  • German

Dataset Structure

gtdh‑hd │ README.md # This file │ classes.json # Class list │ classes_color.json # Class‑to‑color mapping │ classes_discontinuous.json # Class shape information │ classes_ports.json # Electrical port descriptions │ consistency.py # Statistics and consistency checks | loader.py # Simple dataset loading and storage utilities │ segmentation.py # Multi‑class segmentation generation │ utils.py # Helper functions │ requirements.txt # Script dependencies └───drafter_D └───annotations # Bounding‑box annotations │ │ CX_DY_PZ.xml │ │ ... └───images # Raw images │ │ CX_DY_PZ.jpg │ │ ... └───instances # Instance‑segmentation polygons │ │ CX_DY_PZ.json │ │ ... └───segmentation # Binary segmentation maps (stroke vs background) │ │ CX_DY_PZ.jpg │ │ ... ...

File Naming Rules

• D is the global drafter identifier
• X is the global circuit identifier (12 circuits per drafter)
• Y is the local diagram identifier (2 diagrams per circuit)
• Z is the local image identifier (4 images per diagram)

Image Files

• Each image is RGB, stored as jpg, jpeg or png (mixed case extensions).

Bounding‑Box Annotations

• Category labels and mapping tables are in classes.json.
• Annotations follow the PASCAL VOC format.
• Every raw image has a corresponding annotation file.

Known Annotation Issues

• C25_D1_P4 truncates a text label
• C27 truncates some texts
• C29_D1_P1 has an extra text
• C31_D2_P4 is missing a text
• C33_D1_P4 is missing a text
• C46_D2_P2 truncates a text

Instance Segmentation

• Each binary segmentation map has a matching instance‑segmentation polygon file in labelme format.

Segmentation Maps

• Binary maps have the same resolution as their images and contain only black‑white pixels representing drawing strokes vs background.

Netlist Files

• Some images include netlist files in ASC format.

Consistency and Statistics

• Scripts are provided for class distribution, bounding‑box size statistics and consistency checks.

• Run scripts as:

  $ python3 consistency.py

  or for a specific drafter:

  $ python3 consistency.py 15

Multi‑Class (Instance) Segmentation Processing

• Scripts are provided to handle new and existing instance segmentation files.

  $ python3 segmentation.py <drafter_id>

  where <command> can be:

  • transform
  • wire
  • keypoint
  • create
  • refine
  • pipeline
  • assign

Dataset Loader

• Loading and writing utilities are included for training.

  from loader import read_dataset

  db_bb = read_dataset() # Load all bounding‑box annotations db_seg = read_dataset(segmentation=True) # Load all polygon annotations db_bb_val = read_dataset(drafter=12) # Load annotations for drafter 12

  len(db_bb) # Number of samples db_bb[5] # Retrieve any sample

  db = read_images(drafter=12) # Returns list of (image, annotation) pairs db = read_snippets(drafter=12) # Returns list of (image, annotation) pairs

Citation

@inproceedings{thoma2021public, title={A Public Ground‑Truth Dataset for Handwritten Circuit Diagram Images}, author={Thoma, Felix and Bayer, Johannes and Li, Yakun and Dengel, Andreas}, booktitle={International Conference on Document Analysis and Recognition}, pages={20--27}, year={2021}, organization={Springer} }