Abstract

Motivated by the abundant availability of user-generated multimedia content, a data augmentation approach that en-hances an initial manually labelled training set with regions from user tagged images is presented. Initially, object detec-tion classifiers are trained using a small number of manually labelled regions as the training set. Then, a set of posi-tive regions is automatically selected from a large number of loosely tagged images, pre-segmented by an automatic seg-mentation algorithm, to enhance the initial training set. In order to overcome the noisy nature of user tagged images and the lack of information about the pixel level annota-tions, the main contribution of this work is the introduction of the visual ambiguity term. Visual ambiguity is caused by the visual similarity of semantically dissimilar concepts with respect to the employed visual representation and analysis system (i.e. segmentation, feature space, classifier) and, in this work, is modelled so that the images where ambiguous concepts co-exist are penalized. Preliminary experimental results show that the employment of visual ambiguity guides the selection process away from the ambiguous images and, as a result, allows for better separation between the targeted true positive and the undesired negative regions.