Image Quality Metrics for Stochastic Rasterization

Abstract

We develop a simple perceptual image quality metric for images resulting from stochastic rasterization. The new metric is based on the frequency selectivity of cortical cells, using ideas derived from existing perceptual metrics and research of the human visual system. Masking is not taken into account in the metric, since it does not have a significant effect in this specific application. The new metric achieves high correlation with results from HDR-VDP2 while being conceptually simple and accurately reflecting smaller quality differences than the existing metrics. In addition to HDR-VDP2, measurement results are compared against MS-SSIM results.

The new metric is applied to a set of images produced with different sampling schemes to provide quantitative information about the relative quality, strengths, and weaknesses of the different sampling schemes. Several purpose-built three-dimensional test scenes are used for this quality analysis in addition to a few widely used natural scenes. The star discrepancy of sampling patterns is found to be correlated to the average perceptual quality, even though discrepancy can not be recommended as the sole method for estimating perceptual quality. A hardware-friendly low-discrepancy sampling scheme achieves generally good results, but the quality difference to simpler per-pixel stratified sampling decreases as the sample count increases.

A comprehensive mathematical model of rendering discrete frames from dynamic 3D scenes is provided as background to the quality analysis.