Exploring a Large Dataset of Educational Videos Using Object Detection Analysis

Abstract

Processing data and analyzing it are very important these days and have many applications in real life. In the case of very large video datasets, it became counterproductive, if not impossible, to visualize, analyze, and label the videos manually, due to their sheer number and length. Thus, automatically explore video datasets to understand what the videos are depicting, both for having a general view of the videos and a per video understanding of the activities. For example, if they contain sensitive images, violence and other such activities the videos could be flagged for manual checking. We turned our attention toward understanding educational videos, due to the great potential of improving the educational process by understanding written data (such as forums, chats, documents), audio, and video data. In this paper, we continue our work regarding an exploratory analysis of videos collected using YouTube-8 M, with the “school” keyword in their metadata. Thus, we attempt to detect the type of activity in a video based on the number of unique people detected and tracked, and on the objects detected by YOLOv3. This allows us to detect the educational activities in the dataset and to estimate the distribution of the activities in videos uploaded worldwide.