The Automated Generation and Further Application of Tree-Structure Outline for Lecture Videos with Synchronized Slides

Abstract

In this paper, the authors illustrate their motivation and method in the automated generation of tree-structure outline for lecture videos with supplementary synchronized slides, and then propose a further application, lecture video segmentation by slide-group-change event, based on the outline previously generated. Starting with OCR (Optical Character Recognition) result, with an approximate accuracy of 90%, the authors attempt to reconstruct the text system of each slide into an up-to-3-level content tree, and then explore logical relations between slides in order to set them hierarchical. A final up-to-6-level outline will be achieved after removing all the redundancy. And the hierarchy of the slides, which is saved in the outline, will largely simplify the additional segmentation process. Evaluation result shows that, the final outline generated based on the test dataset retains about only 1/4 of the original texts from all slides and is organized well, with a high accuracy of 85% at slide title level. And the majority of the segments the authors' get are logically reasonable, while the average length of them is about 5~15 minutes.