Abstract
Video interpolation is aiming to generate intermediate sequence between two frames. While most existing studies require the two reference frames to be consecutive, we propose a stochastic learning frame work that can infer a possible intermediate sequence between a long interval. Therefore, our work expands the usability of video interpolation in applications such as video long-term temporal super-resolution, missing frames repair and motion dynamic inference. Our model includes a deterministic estimation to guarantee the spatial and temporal coherency among the generated frames and a stochastic mechanism to sample sequences from possible realities. Like the studies of stochastic video prediction, our generated sequences are both sharp and varied. In addition, most of the motions are realistic and can smoothly transition from the referred start frame to the end frame.
Abstract (translated)
视频插值旨在生成两帧之间的中间序列。虽然大多数现有研究要求两个参考框架是连续的,但我们提出了一个随机学习框架工作,可以推断长间隔之间可能的中间序列。因此,我们的工作扩展了视频插值在视频长期时间超分辨率,丢帧修复和动态动态推理等应用中的可用性。我们的模型包括确定性估计,以保证生成的帧之间的空间和时间一致性,以及从可能的现实中采样序列的随机机制。与随机视频预测的研究一样,我们生成的序列既清晰又多变。另外,大多数运动是逼真的并且可以从引用的起始帧平滑过渡到结束帧。
URL
https://arxiv.org/abs/1809.00263