OVERVIEW

In the last few years, machine learning has experienced an astonishing development in all domains. It has become the technique to-go for most of existing image analysis problems reporting state-of-the-art results in tasks such as classification, detection and segmentation. The superior performance is reported in supervised learning, however, this paradigm relies on the strong assumption of having a well-representative annotated dataset. In several real world problems such as image segmentation this assumption is not practical, and the bias in the annotations adversely affect the segmentation outputs. We then motivate the use of fewer labels for segmenting several types of imaging data. In particular, we study the advances of recent unsupervised and semi-supervised methods that exploit large sets of unlabelled data for segmentation. We will discuss the current state-of-the-art techniques, challenges and opportunities.