Automatic vectorization of historical maps: A benchmark.

Area of Science:

• Computer Vision
• Digital Humanities
• Geographic Information Systems (GIS)

Background:

• Historical map digitization is crucial for urban studies, but existing shape vectorization methods face accuracy challenges.
• Digitizing 19th and early 20th-century Paris atlases requires robust techniques for extracting complex geographic details like buildings and streets.

Purpose of the Study:

• To comprehensively evaluate and improve supervised pipelines for accurate and complete shape vectorization of historical maps.
• To identify the most effective methodological choices for historical map digitization and feature extraction.

Main Methods:

• Developed a supervised pipeline combining deep edge filtering and watershed transform for closed shape extraction.
• Proposed an improved training protocol and joint optimization for edge detection and shape extraction stages.
• Compared state-of-the-art deep edge filters (including vision transformers) and evaluated deep learnable watershed against traditional methods.

Main Results:

• Established a benchmark for historical map vectorization, making data, code, and results publicly available.
• Demonstrated the effectiveness of a joint optimization approach for enhancing vectorization performance.
• Identified critical paths for fully automatic extraction of key historical map elements.

Conclusions:

• The proposed pipeline and optimized methods significantly improve the accuracy and completeness of historical map vectorization.
• This work provides a valuable resource for researchers in digital humanities and GIS, enabling new historical analyses.
• The open availability of resources facilitates further research and application in historical cartography and urban morphology studies.