AI

Quantitative evaluation of omnidirectional video quality

Abstract

Omnidirectional video encoding and delivery are rapidly evolving fields, where choosing an efficient representation for storage and transmission of pixel data is critical. Given that there are a number of projections (pixel representations), a projection independent measure is needed to evaluate the merits of different options. We present a technique to evaluate projection quality by rendering virtual views and use this to evaluate three projections in common use: Equirectangular, Cubemap, and Equi-Angular Cubemap. Through evaluation on dozens of videos, our metrics rank the projection types consistently with pixel density computations and small scale user studies.