Domain adaptation and continual learning in semantic segmentation

Deep networks produce outstanding results in many computer vision tasks including semantic segmentation. On the other hand, they are plagued by the need of a huge amount of labeled data for training, which is not always available or may not be accessible all together as typically required in the standard supervised machine learning setting. These problems raise the demand for knowledge transfer techniques able to adapt the learning performed on one domain to a related one and to allow the training of the network in multiple stages. This chapter will start by introducing the domain adaptation task for semantic segmentation and the different levels at which the adaptation can be performed. Then, the different families of strategies will be discussed in detail presenting the most successful approaches for each of them. In the second part, we will present the task of continual learning in semantic segmentation. Although being a relatively new research field, its interest is rapidly growing, and many different scenarios have been introduced, which will be described in detail along with the approaches to tackle them.