Coding Gain and Tuning for Parametrized Visual Quality Metrics

A visual quality metric for video provides a mathematical expression for visual quality. We show how the coding gain of rate-distortion optimal (RDopt) encoding of the quantizer step size depends on the exact shape of this expression. For the Mean Square Error, we find a small coding gain (1% on average) compared to using a constant quantizer. However, we do find a significant gain for the Structural Similarity (SSIM) (12% on average). Finally, we show parameter tuning using the new Perceptually Weighted Error (PWE). The tuning results in a parameter setting of PWE that offers the best visual quality of RDopt encoding.