Synthetic to Real World Image Translation Using Generative Adversarial Networks

In this paper we perform synthetic to real world image translation using Cycle Consistent Adversarial Networks and further provide insights as to how the translation can be utilized to build a framework that can solve a plethora of domain adaptation tasks by reducing the same to a supervised problem via image generation and label transfer from a source domain (synthetic images) to a target domain (real world images) thereby eliminating the need for a labeled target domain dataset. We run experiments taking the (GTA) dataset as the source domain and the Cityscape dataset as the target domain and provide a comprehensive evaluation of the experiments. We also provide analysis and visualization of existing domain adaptation techniques and evaluate their pros and cons. A potential application of the concept is in training of self-driving cars as it is difficult to obtain an extensive labeled dataset for every geographical region.