RoadRouter: Multi-Task Learning of Road Network Extraction with Graph Representation

Abstract

Since most state-of-the-art road mappers pose the road network extraction as a binary segmentation trained on the RGB dataset, our proposed ‘RoadRouter’ system pushes the frontier by classifying the roads into seven categories based on the SpaceNet annotations. Our system is built with the Red/Near-Infrared dataset, making use of the asphalt’s spectral signature to differentiate roads from other influential noises. For addressing the disconnected road gaps problem, we propose the stacked hourglass network with dual supervision. Inspired by the human behavior of tracing the road networks via a constant orientation, incorporating the orientation learning as auxiliary loss leads to more robust and synergistic representations favorable for road connectivity refinement. The intermediate supervision provided by stacking the hourglass modules successively also serves as a connectivity refinement mechanism. In the case of modeling the long-range interaction among the per-pixel predictions, the traditional color-based appearance kernel is not useful in CRF post-processing. We propose the pixel-wise orientation CRF specific for bridging the fragmented road segments. We also formalize an image transformation protocol to parse the topology from the road segmentation. The undirected closed graphs can thereby be constructed from probabilistic inferences. Various graph-based algorithms, e.g., the shortest path searching, can be implemented on the road graph representations.