Abstract
Large motion poses a critical challenge in Video Frame Interpolation (VFI)task. Existing methods are often constrained by limited receptive fields,resulting in sub-optimal performance when handling scenarios with large motion.In this paper, we introduce a new pipeline for VFI, which can effectivelyintegrate global-level information to alleviate issues associated with largemotion. Specifically, we first estimate a pair of initial intermediate flowsusing a high-resolution feature map for extracting local details. Then, weincorporate a sparse global matching branch to compensate for flow estimation,which consists of identifying flaws in initial flows and generating sparse flowcompensation with a global receptive field. Finally, we adaptively merge theinitial flow estimation with global flow compensation, yielding a more accurateintermediate flow. To evaluate the effectiveness of our method in handlinglarge motion, we carefully curate a more challenging subset from commonly usedbenchmarks. Our method demonstrates the state-of-the-art performance on theseVFI subsets with large motion.