Seg2Vector: Remote Sensing Road Graph Extraction via Segmentation-to-Vector Transformation

Road graph extraction from high-resolution remote sensing images is a pivotal task for applications such as urban planning and autonomous driving. However, segmentation-based approaches often suffer from a complex entanglement between semantic feature learning and topological structure modeling. This coupling can result in a misalignment between training objectives and the requirements of inference. To address these limitations, we propose Seg2Vector, a novel framework designed to directly transform segmentation masks into road graphs. Instead of depending on internal deep features extracted from the segmentation model, our method introduces a dedicated vector model that learns geometric representations directly from the segmentation masks. This design effectively decouples the topological reasoning task from the segmentation process. Moreover, unlike prior methods that explicitly predict connectivity between vertices—which can be ambiguous and challenging—the vector model predicts directional vectors for each vertex. These vectors are then processed by a heuristic algorithm to construct the final road graph. Seg2Vector allows the segmentation model and the vector model to be trained independently, enabling a plug-and-play capability across diverse datasets. Extensive experiments on three public benchmarks validate the effectiveness of our method, demonstrating superior performance, particularly in terms of topological correctness.