Representing Feature Location Uncertainties in Spherical Images

Pose uncertainty estimation of calibrated cameras is a common task in the field of computer vision and uses location uncertainties of image features. For spherical cameras, those uncertainties cannot be optimally described using conventional latitude-longitude representation. Increasing distortions close to the poles of the spherical coordinate system prevent a suitable description through Gaussians. To overcome this limitation, we present a consistent location uncertainty representation for spherical image features: Our approach is based on normal vectors in Cartesian space and applicable to any kind of camera with convex projection surfaces, such as catadioptric and spherical systems. We compare its performance against latitude-longitude representation by estimating camera pose uncertainties through first order error propagation in a weighted least squares pose estimation scenario. Our experiments on synthetic and real data show that the proposed approach delivers consistent results outperforming conventional latitude-longitude representation.