Compressed domain video processing of meetings for activity estimation in dominance classification and slide transition detection

Compressed domain processing of video has been a widely used tool in enabling computationally efficient video analysis in the last decade or so since the standardization of video compression in the form of MPEGx/H.26x. We consider the use of such features in the meeting domain to reduce the processing time of video analysis. We review the various compressed domain features that can be extracted easily from compressed videos. In this report, two applications of interest in meeting analysis are considered. First, we present work on activity level estimation which is used in dominance modeling of meeting participants. Second, we look at the problem of detecting slide transitions in meetings, which can be used as contextual cues for estimating the visual focus of attention of meeting participants. In both applications, our experimental results show that compressed-domain methods do as well as their corresponding pixel-domain methods, but only require a fraction of computational costs.