Key slide-segments annotation for presentation videos

This paper presents a bimodal approach to presentation video annotation suitable for supporting content-based search. Our aim is to segment a presentation video into a sum of key slide segments of more manageable length and to annotate these segments by using text derived from the presentation file. In operation, a presentation video is first parsed into perceptually distinct slide segments. A key frame is then extracted from each segment. Slide region in the key frame is then compared against slides in the presentation file by means of template matching. Matched segments satisfying a duration threshold are declared as key slide segments. Non key slide segment are then suppressed and appended to adjacent key slide segments. A rounding step is then applied to ensure match between slide segments and slides in the presentation file. Matched key slide segments are then annotated with text from corresponding slides while segments containing question answer events are identified and tagged by using a speaker turn detection procedure.