Framework for personalized multimedia summarization

  • Authors:

  • Lalitha Agnihotri;John Kender;Nevenka Dimitrova;John Zimmerman

  • Affiliations:

  • Columbia University, New York, NY;Columbia University, New York, NY;Philips Research, Briarcliff Manor, NY;Carnegie Mellon University

  • Venue:
  • Proceedings of the 7th ACM SIGMM international workshop on Multimedia information retrieval
  • Year:
  • 2005

Quantified Score

Hi-index 0.00

Visualization

Abstract

We extend and validate methods of personalization to the domain of automatically created multimedia summaries. Based on a previously performed user study of 59 people we derived a mapping of personality profile information to preferred multimedia features. This article describes our summarization algorithm. We define constraints for automatic summary generation. Summaries should consist of contiguous segments of full shots, with duration proportional to the log of video length, selected by an objective function of total "importance" of features, with heuristic rules for deciding the "best" combination of length and importance. We validated the summaries with a user study of 32 people. They were asked to answer a shortened series of personality queries. Using this current user profile, together with the earlier genre-specific reduced mapping and with automatically derived features, we automatically generated two summaries for each video: one optimally matched, and one matched to the "opposite" personality. Each user evaluated both summaries on a preference scale for four each of: news, talk show, and music videos. From a statistical analysis we find statistically significant evidence of the effectiveness of personalization on news and music videos, with no evidence of user subpopulations. We conclude for these genres that our claim, of a universal mapping from certain measured personality traits to the computable creation of preferred multimedia summaries, is supported.