Live Poster Session:
Thursday, July 29th 2:45-3:45pm EDT
Abstract: In order to better study political messaging, political scientists must be able to extract and analyze the text from campaign advertisements. Currently, popular text extraction tools are too costly to perform on hundreds or thousands of videos. The purpose of this research project is to develop a cost-effective method for video analysis that is as accurate and efficient as the more costly products. To this end, a method was needed to identify a handful of key frames that could be collected to reflect all of the text that is displayed throughout the video. These key frames were identified by counting the number of corners in each frame and calculating the proportion of each frame covered by text. It was found that these key frames occurred at the local maximas of the corner count metric. By stacking the key frames into one image, the text of the video can be represented by just one image. This method of identifying key frames and compacting the video content into one image offers a significant benefit because APIs that offer text extraction for videos charge about ten cents for every minute of video but only a tenth of a cent to process an image. The approach developed here would mean a cost reduction by a factor of 100.
Video: