berboost.blogg.se

On the other hand, the content-dependent results clearly indicate that even perfect (ranking) prediction is reachable for some contents (ie, the results of PU-SSIM for content “LasVegasStore” and “RedwoodSunset”). On one hand, the codec-dependent results show that all the metrics suffer to some extent from content dependency in their prediction capability. In terms of content dependency and distortion dependency, these results confirm widely known observations concerning the scope of validity of most objective metrics ( Huynh-Thu and Ghanbari, 2008).

Error bars indicate 95% confidence intervals. For comparison, results for HDR-VDP-2.1 and HDR-VDP-2.2 are also shown. Spearman correlation coefficients for LDR metrics. Interestingly, PU-SSIM has performance almost equivalent to that of HDR-VDP-2.2, albeit at lower computational cost.įigure 17.6. We also report results for HDR-VDP-2.1 and HDR-VDP-2.2, with the latter improving significantly over the older version of the metric, as discussed in the previous section.

As can be seen, overall the best performing LDR metric is PU-SSIM, followed by log SSIM. 17.6 summarizes the performance of the metrics and shows also the 95% confidence interval of the Spearman correlation coefficient (modulus). The values of the modulus of the correlation coefficient for each metric, per content and per codec, are reported in Table 17.1. The use of a nonparametric correlation coefficient avoids the need for nonlinear fitting to linearize the values of the objective metrics, which may be questionable because of the relatively small size of the subjective ground truth dataset.

The performance of the metrics was evaluated in terms of the Spearman rank-order correlation coefficient, computed on the entire set of mean opinion scores.