Robust Video Content Analysis via Transductive Learning

摘要: Reliable video content analysis is an essential prerequisite for effective search. An important current research question how to develop robust methods that produce satisfactory results a large variety of sources, distribution platforms, genres, and content. The work presented in this article exploits the observation appearance objects events often related particular sequence, episode, program, or broadcast. This motivates our idea considering task single episode as transductive setting: final classification model must be optimal given only, not general, expected inductive learning. For purpose, unlabeled test data have used learning process. In article, framework based on feature selection ensemble presented. contrast approaches (e.g., concept detection), designed general manner only task. proposed applied following tasks: shot boundary detection, face recognition, semantic retrieval, indexing computer game sequences. Experimental diverse tasks sets demonstrate improves robustness underlying state-of-the-art approaches, whereas support vector machines do solve manner.