Iden%fica%on of Narra%ve Peaks in Clips: Text Features Perform Best Joep J.M. Kierkels, Mohammad Soleymani, Thierry Pun Computer Vision and Mul2media Laboratory (h8p:// cvml.unige.ch) Computer Science Department University of Geneva Switzerland 15‐10‐2009 CLEF 2009, Corfu 1
Outline • Mo2va2on • Contest and task • Methods • Results • Conclusion and future work 15‐10‐2009 CLEF 2009, Corfu 2
Mo2va2on Iden2fica2on of narra2ve peaks or drama2c tension moments For Video summariza2on and highligh2ng 15‐10‐2009 CLEF 2009, Corfu 3
Task No narrative peak Narrative peak 15‐10‐2009 CLEF 2009, Corfu 4
Task No narrative peak Narrative peak 15‐10‐2009 CLEF 2009, Corfu 5
Video features • Frame absolute difference (grayscale) • Thresholding of the difference score • Smoothed over 10 seconds long window 15‐10‐2009 CLEF 2009, Corfu 6
Audio features • Pitch increase • Neglec2ng the non‐ increasing samples • Audio Energy (loudness) • Smoothed over 10 second long window 15‐10‐2009 CLEF 2009, Corfu 7
Text features • The introduc2on of a new topic • Removing the rare and frequent words • Genera2on of the vocabulary vector v 15‐10‐2009 CLEF 2009, Corfu 8
Narra2ve peak distribu2on 15‐10‐2009 CLEF 2009, Corfu 9
Fusion • Drama2c tension detec2on based on • The first three peaks with the distance more than 5 seconds were chosen 15‐10‐2009 CLEF 2009, Corfu 10
Fusion and train set results Scheme Features Weights Scheme Features Weights 1 Video Yes 5 Video, Text Yes 2 Audio Yes 6 Audio, Text Yes 3 Text Yes 7 Video, Audio, Yes Text 4 Video, Yes 8 Text No Audio BG_36941 BG_37007 BG_37016 BG_37036 BG_37111 Total Scheme number 1 0 0 1 1 1 3 2 2 1 1 1 1 6 3 2 1 1 2 1 7 4 0 1 2 1 1 5 5 1 2 2 1 0 6 6 2 1 1 2 1 7 7 1 1 2 1 0 5 8 0 1 1 1 0 3 15‐10‐2009 CLEF 2009, Corfu 11
Results on evalua2on set run number (scheme nr) Score (Peak- Score (Point- based) based) 1 3 33 39 2 7 30 41 3 6 33 42 4 8 32 43 5 Random 32 43 15‐10‐2009 CLEF 2009, Corfu 12
Conclusion and future work • Challenging an difficult task! • Text features and seman2cs • Having all the peaks will improve the methods • Using facial expressions • Pitch based detec2on with all the drama2c tension moments • Larger training set, supervised learning 15‐10‐2009 CLEF 2009, Corfu 13
Machine learning protest 15‐10‐2009 CLEF 2009, Corfu 14
Thank you for your a8en2on! Ques2ons, comments, sugges2ons 15‐10‐2009 CLEF 2009, Corfu 15
Recommend
More recommend