Video Frame Interpolation Based on Multi-scale Convolutional Network and Adversarial Training

Abstract

We propose a deep multi-scale convolutional neural network solution for video frame interpolation, which can synthesize the interpolated frames with favorable quality and visual experience. To get sharp results, we use a combination of loss function, including a Wasserstein generative adversarial network loss with gradient penalty. We try a slim generator network structure in order to meet the real-time interpolation requirement as much as possible. In this way our framework contains less parameters, which could be beneficial to video processing tasks in future works. Our work is also shown to be effective in improving subjective visual experience for video frames in most cases.