An Analytical Study of CNN-based Video Frame Interpolation Techniques

Videos are made up of a series of continuous image frames. Given these consecutive two frames, video frame interpolation techniques aim for synthesizing video frame or frames that lie temporally in between the given frames, and that is/are spatially adjusted. Video frame interpolation finds various use cases in computer vision activities including but not limited to video restoration i.e., to generate clear video frames in sections where a video is blurred, generating video animations (software editing tools), view synthesis and so on. Relying upon standard optical flow-based video frame interpolation methods is difficult as they often produce blurry results in the cases of occlusion and extensive motion handling between the objects in the frames. Remarkable achievements have been made in the recent past to capture varied large scale motion using deep convolution networks. In this paper, we will discuss how some deep convolution networks based methods have evolved over the years to improve the quality of the synthesized frames both qualitatively and quantitatively for video frame interpolation task.