On broadcasted game video analysis: event detection, highlight detection, and highlight forecast

Efficient access to broadcasted computer game videos is urgently demanded due to the emergence of live streaming platforms. The popularity of game video streaming builds a big market showing commercial potentials and arising many technical challenges. In this work we facilitate efficient access from two aspects: event detection and highlight detection. By recognizing designated text displayed on screen when important events occur, we associate game events with time stamps, and accordingly develop an interface to facilitate direct access. For highlight detection, we jointly consider visual features, event features, and viewer’s behavior to construct two highlight models based on the psychophysiological approach and the data-driven approach, respectively. The concatenated highlights then enable compact game video presentation. To facilitate adaptive live streaming, a novel highlight forecast model is built to predict whether there will be a highlight in the next seconds, so that the streaming system can allocate more resource for more important segments on the fly. Comprehensive experiments based on various experimental settings demonstrate effectiveness of the proposed methods. We believe that this work is one of the early attempts on analyzing broadcasted computer game videos from the perspective of multimedia content analysis.