Object-level detection of hand-drawn annotations in participatory sketch maps using paired clean and annotated basemaps
Automatic extraction of hand-drawn annotations from participatory sketch maps is essential for digitising community-generated spatial information but remains challenging due to heterogeneous drawing styles, scanning artefacts, and complex basemap content. Existing approaches typically treat markup e...
Gespeichert in:
| Hauptverfasser: | , , , , , |
|---|---|
| Dokumenttyp: | Article (Journal) |
| Sprache: | Englisch |
| Veröffentlicht: |
10 Jun 2026
|
| In: |
AGILE: GIScience series
Year: 2026, Jahrgang: 7, Pages: 1-7 |
| DOI: | 10.5194/agile-giss-7-32-2026 |
| Online-Zugang: | Verlag, kostenfrei, Volltext: https://doi.org/10.5194/agile-giss-7-32-2026 Verlag, kostenfrei, Volltext: https://agile-giss.copernicus.org/articles/7/32/2026/ 
|
| Verfasserangaben: | Clemens Langer, Celina Thomé, Nir Fulman, Steffen Knoblauch, Alexander Zipf, and Yulia Grinblat |
| Zusammenfassung: | Automatic extraction of hand-drawn annotations from participatory sketch maps is essential for digitising community-generated spatial information but remains challenging due to heterogeneous drawing styles, scanning artefacts, and complex basemap content. Existing approaches typically treat markup extraction as pixel-level segmentation or simple image differencing, which struggles under real-world variability. To address this, we formulate annotation extraction as an object-level task using a YOLO-based detector applied to RGB images of annotated maps. In addition, change detection is performed using paired RGB images of annotated and clean maps to isolate user-drawn content from the underlying basemap. Experiments on ∼2,300 real sketch maps and ∼18,000 synthetic samples show strong performance across diverse conditions. Object detection on annotated maps alone achieves mAP50 of 91.5% on satellite imagery and 97.3% on OSM basemaps, while incorporating paired clean maps for change detection improves performance to 97.4% and 98.1%, respectively. Synthetic pretraining further enhances results on real hand-drawn data, indicating that simulated annotations effectively supplement limited labelled samples. |
|---|---|
| Beschreibung: | Gesehen am 12.08.2026 |
| Beschreibung: | Online Resource |
| DOI: | 10.5194/agile-giss-7-32-2026 |