Cinemagraph generation from a static image with generative adversarial networks

Abstract

Cinemagraph is a sequence of still frames which one of the entities in the scene has small movements. The process of making cinemagraph from video has been studied extensively and there are many methods and commercial software available for this problem. However, there are very few research which address the problem of cinemagraph generation from one single static image. In addition, the manual process of generating cinemagraphs from one single static image is tedious and time-consuming. Therefore, automation of this process is necessary. In this research, we presented a novel method based on generative adversarial networks to automate the process of cinemagraph generation from single static image. Our method consists of a two-stage generator and a discriminator. The generator learns foreground, which contains the motion and a mask on how to combine the learned foreground with the input image as the background. The discriminator is responsible for identifying the real data and fake data, which is produced by the generator. Our method achieves the results superior than the state-of-the-art.