Generative Adversarial Networks Based Scene Generation on Indian Driving Dataset

Abstract

The rate of advancement in the field of artificial intelligence (AI) has drastically increased over the past twenty years or so. From AI models that can classify every object in an image to realistic chatbots, the signs of progress can be found in all fields. This work focused on tackling a relatively new problem in the current scenario-generative capabilities of AI. While the classification and prediction models have matured and entered the mass market across the globe, generation through AI is still in its initial stages. Generative tasks consist of an AI model learning the features of a given input and using these learned values to generate completely new output values that were not originally part of the input dataset. The most common input type given to generative models are images. The most popular architectures for generative models are autoencoders and generative adversarial networks (GANs). Our study aimed to use GANs to generate realistic images from a purely semantic representation of a scene. While our model can be used on any kind of scene, we used the Indian Driving Dataset to train our model. Through this work, we could arrive at answers to the following questions: (1) the scope of GANs in interpreting and understanding textures and variables in complex scenes; (2) the application of such a model in the field of gaming and virtual reality; (3) the possible impact of generating realistic deep fakes on society.