An adversarial approach to improve instance segmentation in automotive - Master's Thesis
The purpose of this thesis is to demonstrate how Instance Segmentation can be further improved (allowing in an autonomous driving contect) the vehicle to count on a strong and reliable algorithm, that can help into choosing the right action while driving. To improve such technique, an Adversarial Approach is used, i.e. the Instance Segmentation network is paired with another network, in a Generative Adversarial Network framework, where the two networks are forced to compete and thus to improve each other.
All the details about theory and implementation for this work are described in Thesis PDF document