A Unified Framework for Absolute and Relative Pose Estimation of Spherical Cameras

Spherical cameras are polydiopter cameras that use multiple lenses with overlapping field of view to capture the scene. A spherical view of the scene is reconstructed through stitching algorithms after correcting the distortion introduced by the lenses. These types of cameras are becoming very popular nowadays and are of great interest in several fields ranging from production and distribution of content to virtual reality and autonomous robotics. This paper aims to provide a simple and complete understanding of the spherical camera model by decoupling the model and the equirectangular projection to gain insight into the geometry underlying spherical images. The article assumes that a spherical image has been captured and, delving into the most relevant geometric aspects, proposes algorithms to estimate the pose of a single camera (extrinsic parameters) and the relative position of pairs of cameras (epipolar geometry) without a priori knowledge of the characteristics of the specific device (i.e. optics) used to acquire the image. The article demonstrates that both proposed algorithms can be derived from a common mathematical framework, and the experimental results show that the method for estimating the relative camera pose outperforms the state-of-the-art 8-point algorithm.