Extracting semantic information from news and sport video

Development of systems supporting effective retrieval by content of videos requires the performance of a wide spectrum of operations on video streams, including temporal segmentation, analysis of the audio and video tracks, identification and recognition of text. Low level features are then processed to provide some higher level description of video content, as most of the user queries are typically related to higher level syntax and semantics, rather on the lower, lexical level. Moreover, the specificity of different application domains entails that different solutions be adopted in different contexts. This may affect both the choice of low level features to be extracted as well as the modeling of specific domain knowledge required to address the issue of higher level semantics. In this paper, we report on our experience in the application contexts of news and sports videos. We will show solutions adopted to cope with specific requirements of different application domains