Automatic Evaluation of Video Summaries

摘要

This article describes a method for the automatic evaluation of video summaries based on the training of individual predictors for different quality measures from the TRECVid 2008 BBC Rushes Summarization Task. The obtained results demonstrate that, with a large set of evaluation data, it is possible to train fully automatic evaluation systems based on visual features automatically extracted from the summaries. The proposed approach will enable faster and easier estimation of the results of newly developed abstraction algorithms and the study of which summary characteristics influence their perceived quality.