Geometry Fidelity for Spherical Images

Spherical, or omni-directional, images offer an immersive format appealing to a wide range of computer vision applications. However, the geometric properties of spherical images pose a major challenge for existing models and metrics designed for 2D images. Concretely, we demonstrate that the established generative evaluation metric FID fails to quantify shortcomings in these properties. To this end, we introduce two quantitative evaluation metrics accounting for geometric constraints of spherical images, namely Omnidirectional FID (OmniFID) and Discontinuity Score (DS). OmniFID is an extension of FID, tailored to additionally capture field-of-view requirements of the spherical format by leveraging cubemap projections. DS is a kernel-based seam alignment score of continuity across borders of 2D representations of spherical images. In experiments, OmniFID and DS detect issues with spherical structure better than previously utilized metrics.