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

2026-06-01 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Geospatial AI & Computer Vision · Depth: Expert, quick

Summary

Plan2Map is a new 208-case multimodal benchmark designed for document-grounded geospatial boundary reconstruction using UK planning records. This benchmark requires systems to reconstruct a valid geospatial boundary in GeoJSON format solely from a source planning document, which includes notice text, schedules, map plates, map labels, and boundary annotations. Researchers introduce GeoPlanAgent, a document-grounded, geospatial-tool-in-the-loop system that breaks down the task into evidence extraction, localisation, map registration, boundary segmentation, projection, and verification. GeoPlanAgent achieved a 0.736 mean IoU and 0.904 median IoU on Plan2Map, with 67.8% of predictions at or above 0.8 IoU, significantly surpassing direct VLM-to-GeoJSON baselines. Diagnostic analysis indicates direct VLM prediction remains unreliable, with remaining errors concentrated in localisation and map registration, while supervised boundary segmentation enhances pixel-level mask quality.

Key takeaway

For Computer Vision Engineers developing systems for geospatial boundary reconstruction from planning records, you should prioritize multimodal approaches that decompose the task into distinct stages. Direct VLM-to-GeoJSON methods are unreliable; instead, focus your efforts on improving localisation and map registration components, as these are critical error sources. Consider implementing supervised boundary segmentation to significantly enhance pixel-level mask quality in your solutions.

Key insights

Plan2Map enables multimodal geospatial boundary reconstruction from diverse planning documents via structured processing.

Principles

• Direct VLM prediction is unreliable for geospatial boundaries.
• Decompose complex multimodal tasks for better results.
• Supervised segmentation improves pixel-level mask quality.

Method

GeoPlanAgent reconstructs boundaries by decomposing the task into evidence extraction, localisation, map registration, boundary segmentation, projection, and verification, integrating geospatial tools.

In practice

• Integrate diverse document elements for geospatial tasks.
• Adopt a multi-stage pipeline for document-grounded reconstruction.
• Prioritize improving localisation and map registration steps.

Topics

• Plan2Map Benchmark
• Geospatial Boundary Reconstruction
• Multimodal AI
• Document Analysis
• Computer Vision
• GeoJSON

Best for: AI Scientist, Computer Vision Engineer, Research Scientist

Related on AIssential

• Paper Digest: CVPR 2026 Papers & Highlights — CVPR 2026 highlights reveal advancements in multimodal AI, 3D reconstruction, and generative models.
• From Boundaries to Semantics: Prompt-Guided Multi-Task Learning for Petrographic Thin-section Segmentation — Petro-SAM enables high-quality joint grain-edge and lithology segmentation in petrographic images by adapting SAM.
• Grid2Matrix: Revealing Digital Agnosia in Vision-Language Models — VLMs exhibit "Digital Agnosia," failing to express fine-grained visual details despite their vision encoders retaining the information.
• MAPRPose: Mask-Aware Proposal and Amodal Refinement for Multi-Object 6D Pose Estimation — MAPRPose enhances 6D pose estimation by combining mask-aware proposals with amodal refinement for robust occlusion handling.
• Modeling Depth Ambiguity: A Mixture-Density Representation for Flying-Point-Free Depth Estimation — Modeling depth ambiguity with multiple hypotheses per pixel prevents flying points and improves boundary reconstruction in depth estimation.
• Geometry-Aware Cross Modal Alignment for Light Field-LiDAR Semantic Segmentation — Integrating light field images and LiDAR point clouds significantly enhances semantic segmentation, especially for occluded objects.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.