Scaling up an Unsupervised Image-to-Image Translation Framework from Basic to Complex Scenes

Abstract

Unsupervised image-to-image translation methods have received a lot of attention in the last few years. Multiple techniques emerged to tackle the initial challenge from different perspectives. Some focus on learning as much as possible from the target-style using several images of that style for each translation while others make use of object detection in order to produce more realistic results on content-rich scenes. In this paper, we explore multiple frameworks that rely on different paradigms and assess how one of these that has initially been developed for single object translation performs on more diverse and content-rich images. Our work is based on an already existing framework. We explore its versatility by training it with a more diverse dataset than the one it was designed and tuned for. This helps understanding how such methods behave beyond their original application. We explore how to make the most out of the datasets despite our computational power limitations. We present a way to extend a dataset by passing it through an object detector. The latter provides us with new and diverse dataset classes. Moreover, we propose a way to adapt the framework in order to leverage the power of object detection by integrating it in the architecture as one can see in other methods.