Abstract
Tremendous advances have been achieved in image translation with the employment of generative adversarial networks (GANs). With respect to video-to-video translation, similar idea has been leveraged by various researches, which may focus on the associations among relevant frames. However, the existing video-synthesis methods based on GANs do not make full exploitation of the spatial–temporal information in videos, especially in the continuous frames. In this paper, we propose an efficient method to conduct video translation that can preserve the frame modification trends in sequential frames of the original video and smooth the variations between the generated frames. To constrain the consistency of the mentioned tendency between the generated video and the original one, we propose a tendency-invariant loss to impel further exploitation of spatial-temporal information. Experiments show that our method is able to learn more abundant information of adjacent frames and generate more desirable videos than the baselines, i.e., Recycle-GAN and CycleGAN.