Removing Spinning Motion from Spherical Video Using a Rotationally Invariant Feature Extractor

This paper presents a method for removing rotational motion from an omnidirectional video captured from the viewpoint of a spinning object. For instance, the technique could be used to recover stable footage from spherical video captured from within a football used for competitive sports. The methodology uses a rotationally-invariant algorithm to obtain feature points and descriptors for pairs of successive frames. This information is then abstracted into three-dimensional point clouds, from which the Kabsch algorithm can infer the rotational motion between frames. The resulting rotational matrices are used to remap each equirectangular frame to a reference frame, and thereby offset the effect of frame-to-frame camera rotations. The algorithm has been successfully tested with various styles of footage. Additionally, the codebase implementing our algorithm has been freely shared through an open online repository.