eduzhai > Applied Sciences > Engineering >

MINI-Net Multiple Instance Ranking Network for Video Highlight Detection

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 22 pages

Abstract: We address the weakly supervised video highlight detection problem forlearning to detect segments that are more attractive in training videos giventheir video event label but without expensive supervision of manuallyannotating highlight segments. While manually averting localizing highlightsegments, weakly supervised modeling is challenging, as a video in our dailylife could contain highlight segments with multiple event types, e.g., skiingand surfing. In this work, we propose casting weakly supervised video highlightdetection modeling for a given specific event as a multiple instance rankingnetwork (MINI-Net) learning. We consider each video as a bag of segments, andtherefore, the proposed MINI-Net learns to enforce a higher highlight score fora positive bag that contains highlight segments of a specific event than thosefor negative bags that are irrelevant. In particular, we form a max-max rankingloss to acquire a reliable relative comparison between the most likely positivesegment instance and the hardest negative segment instance. With this max-maxranking loss, our MINI-Net effectively leverages all segment information toacquire a more distinct video feature representation for localizing thehighlight segments of a specific event in a video. The extensive experimentalresults on three challenging public benchmarks clearly validate the efficacy ofour multiple instance ranking approach for solving the problem.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×