Relating Graphical Features with Concept Classes for Automatic News Video Indexing

Automatic indexing of video data, especially news videos, is in strong demand considering their contents' importance and value. Various attempts have been made to index news videos automatically in order to cope with this demand, including recent challenges that utilize accompanying textual information. However, most of these methods tend to be textual information driven, which do not thoroughly consider the image contents. We will propose an indexing method, which considers image contents together with textual information, to ensure the consistency of the video contents and the indexes. This is enabled by rst, acquiring the relations between graphical features and textual concepts from a large volume of training video data. Next, indexing to incoming video is performed by assuming their contents from the acquired relations, referring to the graphical features. In this paper, we will discuss about relating graphical features with concept classes, which is the key technology to enable such indexing.