Marc Yanlong Zhang
Supervisors: Dr. Danda Pani Paudel, Dr. Zhiwu Huang, and Prof. Luc Van Gool
Image to image translation tasks gained popularity over the past years, imposing an interesting challenge in Computer Vision. Data is crucial for those tasks, therefore we utilize weakly paired data to train our models,which are far more common and easier to acquire. The state of the art models are already able to generate very realistic images, the model however is either very large or requires strongly paired data to train. In this thesis we explore the possibility of using weakly paired data to train the model. We use the GPS information of given images to group our data, creating a dataset that does not require much human supervision. Furthermore, we extend the one-to-one translation to a multi domain translation problem. That is, given a discrete set of possible season, time of the day, and weather condition, we are able to translate any image to a specific target domain that is specified by a combination of these three attributes.