Youtubeevent: On Large-Scale Video Event Classification

In this work, we investigate the problem of general event classification from uncontrolled YouTube videos. It is a challenging task due to the number of possible categories and large intra-class variations. On one hand, how to define proper event category labels and how to obtain training samples for these categories need to be explored; on the other hand, it is non-trivial to achieve satisfactory classification performance. To address these problems, a text mining pipeline is first proposed to automatically discover a collection of video event categories. Part-of-Speech (POS) analysis is applied to YouTube video titles and descriptions, and WordNet hierarchy is employed to refine the category selection. This results in 29, 163 video event categories. A POS-based query method is then applied to video titles, and 6, 538, 319 video samples are obtained from YouTube to represent these categories. To improve classification performance, video content-based features are complemented with scores from a set of classifiers, which can be regarded as a type of high-level features. Extensive evaluations demonstrate the effectiveness of the proposed automatic event label mining technique, and our feature fusion scheme shows encouraging classification results.