Adversarial Learning Strategies for Semantic Segmentation

Semantic image segmentation is a computer vision task in which we label specific regions of an image according to their semantic content. This task is of essential importance for a wide range of applications like robotics, autonomous driving, medicine and image editing. Although many datasets have been built for this task, they are typically generic while a specic problem could require to focus more on the data related to it. One of the biggest problems is represented by the difficulty of gathering large datasets. This is caused by the intrinsic complexity and cost of producing fine detailed ground truth for the interested data, as it consists in manually classifying each pixel of the images. In this work we tried to mitigate this problem developing and testing new techniques to perform semi-supervised training and domain adaptation with unlabeled data. Our framework started from some works, presented in the literature, which exploit an adversarial learning framework in order to train a segmentation network using both supervised and unsupervised data. Finally, we developed some extensions that further improve the performances of the unsupervised training process.