Data-Driven Video Completion

by Shachar Ilan · Ariel Shamir

Abstract

Image completion techniques aim to complete selected regions of an image in a natural looking manner with little or no user interaction. Video Completion, the space-time equivalent of the image completion problem, inherits and extends both the difficulties and the solutions of the original 2D problem, but also imposes new ones - mainly temporal coherency and space complexity (videos contain significantly more information than images). Datadriven approaches to completion have been established as a favored choice, especially when large regions have to be filled. In this report we present the current state-of-the-art in data-driven video completion techniques. For unacquainted researchers, we aim to provide a broad yet easy to follow introduction to the subject and early guidance to the challenges ahead. For a versed reader, we offer a comprehensive review of the contemporary techniques, sectioned out by their approaches to key aspects of the problem.