Topology-Aware Neural Vectorization of Orienteering Maps: Contour Segmentation as a Demanding Test Case for Editable Reconstruction

Borbás, Péter and Troll, Ede (2026) Topology-Aware Neural Vectorization of Orienteering Maps: Contour Segmentation as a Demanding Test Case for Editable Reconstruction. In: Proceedings of the 13th International Conference on Applied Informatics. Líceum Kiadó, Eger, pp. 66-80. ISBN 9789634963271

Preview

Text ICAI2026-pp66-80.pdf - Published Version Download (4MB) | Preview

Abstract

We present a reproducible end-to-end pipeline for converting printed orienteering maps into editable vector layers. While the broader framework targets multiple ISOM symbol classes, this paper focuses on elevation contours as a particularly demanding test case because they are central to terrain interpretation, frequently occluded by overprinted symbols, and sensitive to topological degradation during raster-to-vector conversion. To assess practical usefulness rather than raster similarity alone, we combine overlap metrics (Dice, IoU, Boundary-F1) with topology- and structureaware measures, including Betti error, warping error, and graph-level error proxies. The experiments reveal a consistent raster–vector mismatch: models with nearly identical raster validation scores can produce materially different vector graphs after tracing. In particular, U-Net with a ResNet-18 encoder provided the best efficiency–quality trade-off despite near-tied raster scores with heavier variants, while topology-aware losses improved topologyrelated raster metrics without a comparably strong advantage in the postvectorization evaluation. These results show that editable map reconstruction must be evaluated end-to-end in the vector domain, since raster overlap alone is not a reliable proxy for downstream structural fidelity or subsequent editing effort.

Actions (login required)

View Item