SF-Net: Video Frame Interpolation With a 3D Square Funnel Network

Video frame interpolation (VFI) is a problem of designing in-between frames from both the previous and the subsequent frames for enhancing the quality of video. The majority of traditional methods, particularly U-Net-based approaches, suffer from high computational complexity and memory usage in terms of high numbers of parameters. We propose the square funnel network (SF-Net), a novel network structure with significantly fewer parameters but comparable performance, in this paper. SF-Net follows a unique configuration that increases the third dimension of the input frames in deeper layers instead of increasing the number of filters, which results in a more efficient and more compact model. Our model makes use of a maximum of 64 filters in nearly all of its layers, except for the last two layers, which employ 128 filters each. With both objective and subjective evaluation, SF-Net has outstanding visual quality and efficiency, which makes it suitable for low-computational-resource applications. The paper depicts a good direction of VFI, which is to decrease the number of parameters without sacrificing performance.