Plan2Map: A Multimodal Benchmark for Document-Grounded Geospatial Boundary Reconstruction from Planning Records

Fabian Degen, Oishi Deb, Jindong Gu, Junchi Yu, Samuele Marro, Philip Torr, Jialin Yu · 2026

Read Plan2Map: A Multimodal Benchmark for Document-Grounded Geospatial Boundary Reconstruction from Planning Records by Fabian Degen, Oishi Deb, Jindong Gu, Junchi Yu, Samuele Marro, Philip Torr, Jialin Yu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-02747.txt.

Description

Degen and co-authors build a benchmark for drawing maps from planning documents. UK planning records restrict land use over set areas. Their documents rarely give machine-readable boundaries. Evidence is indirect: notice text, schedules, map plates, and labels. Plan2Map holds 208 such cases. Systems must rebuild a valid boundary from the document alone. Reference GeoJSON boundaries are held out for scoring.

Their system, GeoPlanAgent, works with geospatial tools in the loop. It decomposes the task into evidence extraction, localisation, registration, segmentation, projection, and verification. Mean IoU reaches 0.736, with median IoU at 0.904. Two-thirds of predictions score at or above 0.8 IoU. Direct vision-model-to-GeoJSON baselines fall far behind. Error analysis concentrates the remaining faults in localisation and map registration. Supervised boundary segmentation lifts pixel-level mask quality sharply. The benchmark grounds multimodal geospatial work in public records.

Catalog record

CREATORFabian Degen, Oishi Deb, Jindong Gu, Junchi Yu, Samuele Marro, Philip Torr, Jialin Yu
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE79 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.02747

Browse the full library · About the Tower of Babel