Yan Yichao Xu Jingwei Ni Bingbing Yang Xiaokang

Abstract

This work make the first attempt to generate articulated human motionsequence from a single image. On the one hand, we utilize paired inputsincluding human skeleton information as motion embedding and a single humanimage as appearance reference, to generate novel motion frames, based on theconditional GAN infrastructure. On the other hand, a triplet loss is employedto pursue appearance-smoothness between consecutive frames. As the proposedframework is capable of jointly exploiting the image appearance space andarticulated/kinematic motion space, it generates realistic articulated motionsequence, in contrast to most previous video generation methods which yieldblurred motion effects. We test our model on two human action datasetsincluding KTH and Human3.6M, and the proposed framework generates verypromising results on both datasets.