Mapsnap: Automated Georeferencing for Historic Sanborn Insurance Maps

Mapsnap automates georeferencing of historic Sanborn insurance maps, making it far faster to align detailed block-by-block maps from 1870-1960 with modern maps. The project demonstrates that a largely automated pipeline can produce high-quality fits - often accurate enough to line up individual buildings across a century - by extracting street labels, turning them into latitude/longitude control points via OpenStreetMap, and choosing a robust fit. This removes much of the tedious manual work researchers and hobbyists face when locating historic addresses and matching old photographs to vanished streetscapes, and aims to georeference at least 90% of Library of Congress pages without major mistakes.

The pipeline uses EasyOCR (run locally) to detect rotated street labels and returns a street name, a rotated bounding rectangle, and a confidence score; the recognizer was tuned with a city street vocabulary and retrained top layers for Sanborn fonts. Labels are extrapolated into straight-line street directions, intersections are computed and matched to OSM to create candidate ground control points, and a RANSAC-like procedure tests pairs of GCPs to maximize inliers (street distance and angle) and reject noisy detections. The method tolerates large amounts of spurious reads so long as some correct signals exist. Failure modes include short or prefixed names, curved streets, sparse waterfront/rail areas, and renamed streets, which require additional strategies to reach full coverage.