Objective Assessment of Perceived Geometric Distortions in Viewport Rendering of 360° Images
Falah Jabar, João Ascenso, Maria Paula Queluz · IEEE Journal of Selected Topics in Signal Processing · 2019
To render omnidirectional (or 360°) visual content, a projection that maps the pixels from a portion of the viewing sphere to a 2D plane must be employed; this projection creates the viewport image shown to the user and thus has an important role on the quality of experience, for this type of content. However, a sphere to planar projection always introduces geometrical distortions on the rendered image, such as stretching and/or bending of some image regions and structures, which may impact negatively the quality offered to the users. In this article, a content-aware objective quality metric, that predicts the perceived viewport quality, is proposed for the general perspective projection (GPP) rendering of 360° images, which includes the popular rectilinear and stereographic projections. The proposed metric relies on two set of features that characterize the bending of straight lines and stretching of image regions. The extracted features, and the corresponding viewport subjective quality CMOS (comparative mean opinion scores) are then used to build a quality prediction model, based on Support Vector Regression (SVR). The experimental results show that the proposed metric is able to predict the viewport CMOS with a Pearson correlation coefficient close to 0.8. Furthermore, it is shown that the proposed metric may be used to accurately find the GPP projection center that minimizes the perceived geometric distortions, according to the viewport content.