Adapting appearance models of semantic concepts to particular videos via transductive learning
Ralph Ewerth, Bernd Freisleben · 2007
The detection of high-level concepts in video data is an essential processing step of a video retrieval system. The meaning and the appearance of certain events or concepts are strongly related to contextual information. For example, the appearance of semantic concepts, such as e.g. entertainment or news anchors, is determined by the used editing layout which usually is typical for a certain broadcasting station. In recent years, supervised machine learning approaches have been extensively used to learn and detect high-level concepts in video shots. The class of semi-supervised learning methods incorporates unlabeled data in the learning process. Transductive learning is a subclass of semi-supervised learning: In the transductive setting, all training samples are labeled, but the unlabeled test samples are considered in the learning process as well. Up to now, transductive learning has not been applied for the purpose of video indexing and retrieval. In this paper, we propose transductive learning, realized by transductive support vector machines (TSVM), for the detection of those high-level concepts whose appearance is strongly related to a particular video. For each video and each concept, a transductive model is learned separately and adapted to the appearance of a specific concept in the particular test video. Experimental results on TRECVID 2005 video data demonstrate the feasibility of the proposed transductive learning approach for several high-level concepts.