Deep Video Frame Interpolation for Sports Content

Abstract

In this work, we will explore and tackle the problem of video frame interpolation, which aims to artificially generate sequences of frames that are temporally consistant with existing footage. More specifically, we investigate the applications of deep learning methods to the production of high resolution super slow-motion footage.

This thesis is done in collaboration with EVS Broadcast Equipment. We approach the problem around the improvement of their solution, XtraMotion.

We narrow down zones of improvements with probes agnostic to the implementation of XtraMotion and eventually introduce the problem of blinking, our contributions relates to the characteristics, causes and solutions to that specific problem through an extensive analysis of interpolation models.