MGR-Road: Resolving Semantic-Geometric Conflict in One-Shot Vectorized Road Extraction

Abstract

One-shot vectorized road extraction utilizing vision foundation models (VFMs) has established an efficient framework but suffers from a critical semantic-geometric feature conflict. This conflict is reflected in the erosion of high-frequency geometric details by semantic smoothing and negative gradient interference between topology and segmentation optimization. In this letter, we propose multiscale geometric-semantic reconciliation (MGR)-road, a framework designed to resolve this conflict internally across three levels: representation, optimization, and spatial. At the representation level, we implement a collaborative optimization mechanism with layer-wise learning rate decay (LLRD) to preserve low-level geometric features. At the optimization level, we incorporate a soft decoupling mechanism with mask guidance to pass segmentation priors without gradient interference. At the spatial level, a deterministic region-aware sampler is designed to rectify spatial misalignment under curved or occluded geometries. Experiments on CityScale and SpaceNet datasets demonstrate that MGR-Road effectively recovers disconnected topologies, improving topological recall by 1.69% with substantially accelerated convergence compared to the standard segment anything model (SAM)-based baseline. Our approach offers an excellent balance between accuracy and training efficiency, providing a robust solution for resource-constrained applications. The source code is available at https://github.com/sandymoredoit/MGR-Road