HafenCity Trains AI on Fake Historical Maps to Decode Real Ones
Researchers built a pipeline that fakes historical map corpora from OpenStreetMap data, letting a segmentation network hit 87% accuracy with zero real training labels.
ProTakeaways
- New pipeline bootstraps unlimited synthetic historical maps by restyling OpenStreetMap vector data.
- Simulates dust, mildew, and fading via CycleGAN, UNSB diffusion, or manual stochastic degradation.
- SCGCN trained only on synthetic data hits 87% accuracy on real Berlin Straube maps.
- CycleGAN-degraded data beats UNSB downstream despite UNSB having a better FID score.
- Code on GitHub, test set on Zenodo, interactive viewer online.
- Cuts annotation time for a new corpus from weeks of manual labeling to hours.
Historical maps are a goldmine for studying urbanization, land use change, and environmental history, but they resist modern segmentation models. Every corpus has its own quirks: faded ink, dust stains, mildew, idiosyncratic typography. Hand-labeling enough patches to train a network is soul-crushing work. A team from HafenCity University and the University of Bonn just published a method that sidesteps the labeling problem entirely by generating synthetic historical maps from OpenStreetMap data, then training a graph-based segmenter on those fakes.
They call the approach bootstrapping: take modern vector data with known semantics, repaint it in the visual style of a specific historical corpus, and add simulated aging artifacts. Because the vector source already carries land-cover labels, every synthetic map ships with a pixel-perfect ground truth for free.
Why historical maps break normal training pipelines
Supervised segmentation demands large volumes of pre-annotated high-quality data, which barely exists for historical maps, especially when you restrict yourself to a single cartographic style or period. Corpora vary in typeface, symbology, color palette, and scale, and they also carry data-dependent uncertainty from degradation effects such as color fading, folding marks, dust accumulation, or mildew stains. That last bucket, known as aleatoric uncertainty, is what breaks off-the-shelf transfer learning: a model trained on clean modern tiles has never seen a coffee-colored mildew stain smeared across a building footprint.