Answers are grounded in this paper and cite their sources.
Demonstrates effective vector database updates in various geographical datasets, highlighting a novel decision framework.
Keeping vector building databases current requires deciding which database operations should be applied, rather than merely detecting binary geometric differences. Minor discrepancies caused by acquisition noise, representation style, or cartographic generalization may trigger unnecessary updates when using hand-tuned geometric thresholds. We present V-BUILD, a hybrid decision framework that assigns each matched building unit one of four operations: addition, deletion, replacement, or preservation. Deterministic rules handle existential decisions, while a geometry-aware graph Transformer learns the more ambiguous morphological decisions between replacement and preservation directly in the vector domain. The model encodes 1:1 to m:n matching relationships via Unified Correspondence Graphs and uses Adaptive Multi-stream Fusion to integrate node-, edge-, and graph-level geometric descriptors. Experiments on Amsterdam TOP10NL, Hamburg LoD1, and Mercer County OSM data, together with a difficult Amsterdam subset, show that V-BUILD performs well in within-domain evaluations on both authoritative and crowdsourced vector data, while leave-one-domain-out analysis indicates that direct cross-source transfer remains challenging under strong domain shift. On Amsterdam, V-BUILD achieves the best Replacement F1 of 89.44 %; on Mercer, it obtains the best Macro F1 of 88.51 %, outperforming the strongest GNN baseline by 4.22 percentage points; and on the difficult Amsterdam subset, it achieves a 4.51-percentage-point higher Replacement F1 than the strongest GNN baseline, with the associated reduction in paired sample-level errors confirmed by a two-sided exact McNemar test. These results indicate that V-BUILD reduces dependence on manually tuned geometric decision thresholds and provides a practical decision module for incremental vector building database updating.
No takes yet. Share an insight, caveat, or question.
Qiu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: