Video prediction is the challenging task of predicting the future frames of a video, given a sequence of previously observed frames. Although traditional methods have struggled to make progress in this field, it has seen a resurgence of interest in recent years due to the expressive power of deep models. Current video prediction techniques are capable of making near-perfect predictions on synthetic datasets and predicting several steps ahead without diverging on real-world videos. In this paper we discuss the background of video prediction, fundamental problems, applications, and various attempts to solve the problem. We conclude by outlining potential directions for future research.
related paper