Using GANs to Generate Random Surface Materials for High Fidelity Low Altitude Imaging Simulations
Sournav Sekhar Bhattacharya, Zachary Summers, Aditi Panchal, Elise Koock, Neil G. McHenry, Gregory E Chamitoff · 2020
The various emerging technologies surrounding the use of generative machine learning allow for the automation of procedures which would normally take a large amount of human development time. This can be applied to the domain of simulations in low altitude flight imaging simulations. Low altitude imaging consists of images taken at an altitude of less than 500 feet above the surface in this context. Simulating the acquisition of such images using commercial game engines such as Unity or Unreal Engine 4 is possible. However one defect in the images thus taken is the repetition of material images on the terrain itself. This lowers the fidelity of the simulation since materials that represent certain characteristics of the ground are based on a single or small group of source images. In the past few years a set of generative neural network architectures known as generative adversarial networks (GANs) have seen much development. We propose a method using GANs in order to generate variations of similar terrain with a high degree of randomness in order to replace tiling caused by using similar images. We have been sucessful in developing textures for a variety of terrain types including sand and grass by utilizing the evaluation functions and blending algorithms we have constructed. Using these images as terrain textures will allow for a higher fidelity simulation and will provide a more accurate test bed for computer vision algorithms operating in the domain of low altitude flight.