Automatic Spoiler-Sensitive Video Preview Generation With Reinforcement Learning

Abstract

With the plethora of video content uploaded to the internet each day, the amount of watchable content far exceeds the amount of time available for us to consume it. That being said, the aim of this research is not to create something that does the watching for us and merely generates a short summary; instead, the goal is to help us select videos that we would enjoy watching the most while preserving the viewing-integrity of the original video. In other words, we wish to be able to generate preview video summaries that do not reveal any essential plot items. To accomplish this task, we developed an algorithm that is able to select a fraction of a video's total frames to create a video trailer/summary while excluding important frames that would detract from the main viewing experience (i.e. it does not include any main plot points, spoilers, plot twists, etc.). This output video would allow us to decide better whether we are interested in a particular video, and at the same time will still allows us to fully enjoy the video if we do choose to watch it since we will have avoided any spoilers.