Abstract: Early maps of Bohemia from the 16th to 18th centuries record how cartographers saw the river network, but extracting it by hand is slow and every map is drawn differently. Classical image processing struggles to tell rivers from the many other lines on an engraved map, such as letters, hill symbols and ornaments, because it judges shape without context. We show how a transformer-based segmentation network (SegFormer), pretrained on road cracks and fine-tuned on only 260 annotated map tiles, extracts river networks from full map scans. We also cover the problems we had to solve along the way, from a loss for thin structures to vectorization for GIS. The talk closes with a cartographic analysis of how the depicted networks evolved (with Charles University) and first results on colonial maps of Virginia (with the University of Virginia).
Who: Jan Schier, UTIA CAS
When: 10:00 a.m. Friday, October 16
Where: The session will be held in person at the Institute of Information Theory and Automation (UTIA), in room 45 (café). For directions to the institute, please refer to the following link: https://zoi.utia.cas.cz/index.php/contact
Language: Czech



