Video Frame Interpolation via Using GAN to Improve CNN Performance

Video frame interpolation generates intermediate frames between two consecutive ones. Convolutional neural networks (CNNs) achieved good results on computer vision and were frequently used in video frame interpolation. However, in large motion regions, it is challenging to collect enough motion information. The aim of our experiments is to apply generative adversarial networks (GANs) to improve frames generated by CNNs. In this paper, we propose a network that combines CNNs with GANs. We first use CNN as the generator, and then design a discriminator for adversarial learning. Experiments on Vimeo90K and UCF101 demonstrate that our model can attain outstanding performance than the state-of-the-art methods.