Refining inferred road maps using GANs

Abstract

Mapping road networks is both expensive and labor-intensive. A variety of automated mapping approaches have been proposed in recent years, but these schemes often produce maps that are messy, error-prone, or visually unappealing. To fix this, we train a conditional Wasserstein GAN to refine the inferred road map and improve its realism. We show that adding a truth padding stage between the discriminator and generator vastly improves tile consistency, and we introduce a postprocessing pipeline to further clean the graph. To evaluate these results, we focus on a state-of-the-art map inference method known as RoadTracer, published in 2018 by MIT and QCRI. We compare our refinement approach with the original RoadTracer input and easily see qualitative improvements with similar topology.