TUZ at TRECVID 2015: Video Hyperlinking Task

In this paper, we present our video hyperlinking systems for the TRECVID 2015 Video Hyperlinking Task [1]. We used the provided BBC Dataset video keyframes and subtitles to develop two different systems and submit two separate runs. Our first run (tv15lnk TUZ L 1 F M M MERGE1) uses subtitles to discover possible semantic links between video segments. Our second run (tv15lnk TUZ L 1 F M M MERGE2) follows a different approach and uses only visual sentences extracted from keyframes to discover visual links between video segments. When we compare our two different approaches w.r.t. their MAP scores, subtitle based linking performs better. This is probably due to the fact that speech text contains more robust semantic data then visual sentences. We were planning to use both the first and the second systems to get a better result but our third system was not ready in time for submission. Overall results of the TRECVID 2015 Video Hyperlinking Task shows that our subtitle based first system is placed at the third quartile (%50%75) among all participants whereas our visual sentence based second system is placed at the fourth quartile /%75-%100). We are planning to further our work by merging our two systems and discovering more interlinks between BBC Dataset videos.