Abstract
Spherical or omni-directional images offer an immersive visual formatappealing to a wide range of computer vision applications. However, geometricproperties of spherical images pose a major challenge for models and metricsdesigned for ordinary 2D images. Here, we show that direct application ofFr\'echet Inception Distance (FID) is insufficient for quantifying geometricfidelity in spherical images. We introduce two quantitative metrics accountingfor geometric constraints, namely Omnidirectional FID (OmniFID) andDiscontinuity Score (DS). OmniFID is an extension of FID tailored toadditionally capture field-of-view requirements of the spherical format byleveraging cubemap projections. DS is a kernel-based seam alignment score ofcontinuity across borders of 2D representations of spherical images. Inexperiments, OmniFID and DS quantify geometry fidelity issues that areundetected by FID.