Open set object recognition

Abstract

Deep Learning is a widely used technique for classification tasks. In practise, the most common classifiers are not useful for certain tasks, as they were developed to work in an environment were the number of classes is bounded in training phase. In this thesis, we present an alternative classifier that is able to deal with data that belongs to new classes during testing time. This type of data, in which the number of classes is not defined, is referred to as open data. There are some Machine Learning classifiers that have been modified to work with open sets, specially based on SVM. Conversely, in the context of Deep Learning open data is a relatively new area of research. In this thesis we work with OpenMax classifier, showing its improvement when working with open data while also achieving similar results to traditional classifiers for known data.