Inter-active learning of ad-hoc classifiers for video visual analytics

摘要: Learning of classifiers to be used as filters within the analytical reasoning process leads new and aggravates existing challenges. Such are typically trained ad-hoc, with tight time constraints that affect amount quality annotation data and, thus, also users' trust in classifier trained. We approach challenges ad-hoc training by inter-active learning, which extends active learning integrating human experts' background knowledge greater extent. In contrast not only does include expertise posing queries instances for labeling, but it supports users comprehending model visualization. Besides manually or automatically selected instances, empowered directly adjust complex models. Therefore, our visualization facilitates detection correction inconsistencies between examples user's mental class definition. Visual feedback helps assess performance build up filter created. demonstrate capabilities domain video visual analytics compare its results random sampling uncertainty sets.