Video Object Segmentation: Methods and Datasets

Abstract

Video Object Segmentation (VOS) has been one of the several tasks where deep learning has brought enormous performance gains. This task consists in segmenting the objects in a video, that is, grouping together pixels that belong to the same object, both in space (within a frame) and in time (across different frames). Sergi's dissertation focuses on two important aspects to advance the progress in the topic: it proposes new VOS methods that advance the state of the art and it releases new datasets and benchmarks on which such techniques can be trained and compared.