École Nationale Supérieure de
Techniques Avancées

Abstract

Domain adaptation is a field of machine learning enabeling the training of a model on a set of data with its ground-truth labels (synthetic images for instance) that is robust when used on another set of similar but diffrent data (images shot in the real world for example). Training uses both datasets but the ground-truth labels do not exist for the second dataset. We have built a multimodal dataset with infrared, color, and depth images of Paris and its surroundings in order to test multiple domain adaptation methods. At the same time, we worked on a new domain adaptation method which could improve the state of the art in the field of machine learning. This technique is based on the use of a GAN (Generative Adversial Network) to transfer images from the soure domain to the target domain. These images are then mixed up with images from the target domain to train a semantic segmentation model.