kLCRNet: Fast Road Network Extraction via Keypoint-Driven Local Connectivity Exploration

Mingming Zhang, Bin Wang, Shuai Yang, Qingjie Liu, Yun-Hong Wang

Abstract: Road network extraction from remote sensing images has been extensively studied in recent decades. While many approaches output road networks in vector format, most are not fully end-to-end, requiring time consuming postprocessing steps. In addition, challenges like isomorphic encoding limit the flexibility of these methods. In this article, we present kLCRNet, an efficient road network extraction framework that overcomes these limitations by leveraging keypoint-driven local connectivity exploration. kLCRNet consists of two key components: A keypoint detection module that identifies road keypoints via heatmap-based detection and refines them using bipartite matching, and a local connectivity exploration module that samples local connection relationships to directly construct connectivity between detected keypoints. Experiments on the CityScale and SpaceNet datasets demonstrate that kLCRNet outperforms state-of-the-art methods in topological accuracy and connectivity. In addition, kLCRNet significantly improves inference speed by up to 25 times, highlighting its efficiency and effectiveness.