Comparing GANs for Translating Satellite Images to Maps
Arnav Parekhji, Mansi Pandya, Pratik Kanani · 2021
With rapid development taking place in the world, the landscape of the Earth is constantly changing. Skyscrapers now stand in places that were once covered with water. With the advent of continuously evolving technologies, Generative Adversarial Networks are growing at a rapid pace with new variations. This engenders the need to develop and identify an efficient method to accurately convert satellite images into maps so as to adapt to the changes with ease. This is achieved using a conditional Generative Adversarial Network (cGAN) which generates an output image based on certain conditions of the input. Traditionally, image translation is performed using the Pix2Pix GAN which is a type of cGAN. However, the newly developed CycleGAN improves on the problems of the Pix2Pix GAN by making it less selective with respect to the type of data available. The proposed idea is to use both the GANs to perform image to image translation so as to convert satellite images into maps and evaluate their performance using the Frechet Inception Distance. The evaluations are then used to conclude which one is better at generating a model under similar training conditions on the same data. This can help in providing a general overlay of an area within seconds which could be useful to governments in obtaining the layout of a region on which they plan to carry out new projects, or even the armed forces in performing recon of a certain area.