Learning to Take Directions One Step at a Time

摘要: We present a method to generate video sequence given single image. Because items in an image can be animated arbitrarily many different ways, we introduce as control signal of motion strokes. Such also automatically transferred from other videos, e.g., via bounding box tracking. Each stroke provides the direction moving object input and aim train network animation following such directions. To address this task design novel recurrent architecture, which trained easily effectively thanks explicit separation past, future current states. As demonstrate experiments, our proposed architecture is capable generating arbitrary number frames Key components are autoencoding constraint ensure consistency with past generative adversarial scheme that images look realistic temporally smooth. effectiveness approach on MNIST, KTH, Human3.6M, Push Weizmann datasets.