DeepUrbanMapper - Satellite Image to Google Maps Translation

Abstract: The research presents DeepUrbanMapper, an innovative framework developed for the translation of satellite imagery into Google Maps-style renderings utilizing advanced image-to-image translation techniques. It is based on Generative Adversarial Networks (GANs), widely known in machine learning for their ability to yield high-quality images. To train it, the process involves a carefully curated and labeled dataset that consists of paired satellite and map images, enabling DeepUrbanMapper to learn the complex mappings between these two distinct visual domains. The GAN architecture used in DeepUrbanMapper is carefully designed and optimized to preserve spatial coherence and improve visual quality of the translated image. This helps in producing output maps that are close to the ground truth in both visual and geographical aspects. The proposed method includes few novel strategies to stabilize the GAN training and to overcome the notorious mode collapse problem and ensure a consistent output. DeepUrbanMapper has been extensively evaluated quantitatively and qualitatively for its performance. The evaluation shows that the new framework significantly outperforms the existing methods in terms of visual realism of the translated image and its ability to retain detailed features of the input satellite image which opens up many possibilities urban planning, car navigation systems, GIS.