Generative Adversarial Models for Unsupervised Domain Adaptation in Semantic Segmentation

In this thesis we implement an unsupervised domain adaptation framework designed for semantic segmentation and tested in a synthetic-to-real adaptation scenario. First, we propose a pixel-level adaptation strategy based on a cross-domain image mapping to transfer visual attributes from an unlabelled target dataset to a fully annotated source one. In a second moment, we introduce an additional feature-level adaptation to enforce feature distribution alignment between source and target domains.