Local Features, Structure-from-motion and View Synthesis in Spherical Video

This thesis addresses the problem of synthesising new views from spherical video or image sequences. We propose an interest point detector and feature descriptor that allows us to robustly match local features between pairs of spherical images and use this as part of a structure-from-motion pipeline that allows us to estimate camera pose from a spherical video sequence. With pose estimates to hand, we propose methods for view stabilisation and novel viewpoint synthesis. In Chapter 3 we describe our contribution in the area of feature detection and description in spherical images. First, we present a novel representation for spherical images which uses a discrete geodesic grid composed of hexagonal pixels. Second, we extend the BRISK binary descriptor to the sphere, proposing methods for multiscale corner detection, sub-pixel position and sub-octave scale refinement and descriptor construction in the tangent space to the sphere. In Chapter 4 we describe our contributions in the area of spherical structure-from-motion. We revisit problems from multiview geometry in the context of spherical images. We propose methods suited to spherical camera geometry for the spherical-n-point problem and calibrated spherical reconstruction. We introduce a new probabilistic interpretation of spherical structure-from-motion which uses the von Mises-Fisher distribution in spherical feature point positions. This model provides an alternate objective function that we use in bundle adjustment. In Chapter 5 we describe our contributions in the area of view synthesis from spherical images. We exploit the camera pose estimates made by our pipeline and use these in two view synthesis applications. The first is view stabilisation where we remove the effect of viewing direction changes, often present in first person video. Second, we propose a method for synthesising novel viewpoints.