PoSNet: 4x Video Frame Interpolation Using Position-Specific Flow

Abstract

Video frame interpolation has been studied for a long time; however, it is still a difficult low-level vision task. Owing to the improved performance of optical flow estimation, frame-interpolation studies based on optical flow are actively conducted. However, the existing methods are generally tested using high-fps sequences and developed for 2× upscaling or generating multiple frames with a single estimator. This paper proposes a 4× video-interpolation framework that aims to convert 15-fps to 60-fps videos based on a structure comprising flow estimation followed by an enhancement network. We improve the performance by training specialized flow estimators for each direction and frame position. Furthermore, we use the original frames and flow maps as additional inputs for the enhancement network to improve the subjective image quality. Consequently, the proposed network interpolates high-quality frames with a fast runtime and demonstrates its superiority in the AIM 2019 video temporal super-resolution challenge. The associated code is available at https://github.com/SonghyunYu/PoSNet.