Learning Deep Video Stabilization without Optical Flow

  • 2020-11-19 07:26:06
  • Muhammad Kashif Ali, Sangjoon Yu, Tae Hyun Kim
  • 40

Abstract

Learning the necessary high-level reasoning for video stabilization withoutthe help of optical flow has proved to be one of the most challenging tasks inthe field of computer vision. In this work, we present an iterative frameinterpolation strategy to generate a novel dataset that is diverse enough toformulate video stabilization as a supervised learning problem unassisted byoptical flow. A major benefit of treating video stabilization as a pure RGBbased generative task over the conventional optical flow assisted approaches isthe preservation of content and resolution, which is usually obstructed in thelatter approaches. To do so, we provide a new video stabilization dataset andtrain an efficient network that can produce competitive stabilization resultsin a fraction of the time taken to do the same with the recent iterative frameinterpolation schema. Our method provides qualitatively and quantitativelybetter results than those generated through state-of-the-art videostabilization methods. To the best of our knowledge, this is the only work thatdemonstrates the importance of perspective in formulating video stabilizationas a deep learning problem instead of replacing it with an inter-frame motionmeasure

 

Quick Read (beta)

loading the full paper ...