Techniques for Evaluating Objective Video Quality Models Using Overlapping Subjective Data Sets

This report presents techniques for evaluating objective video quality models using overlapping subjective data sets. The techniques are demonstrated using data from the Video Quality Experts Group (VQEG) Multi-Media (MM) Phase I experiments. These results also provide a supplemental analysis of the performance achieved by the objective models that were submitted to the MM Phase I experiments. The analysis presented herein uses the subjective scores from the common set of video clips to map all the subjective scores from the 13 or 14 experiments (at a given image resolution) onto a single subjective scale. This mapping greatly increases the available data and thus allows for more powerful analysis techniques to be performed. Resolving power values are presented for each model and image resolution. On a per-clip level, models' responses to stimuli are analyzed with respect to all stimuli, each coding algorithm, coding-only impairments, and transmission error impairments. The models' responses to stimuli are also analyzed on per-system and per-scene levels. Results indicate the amount of improvement possible when averaging over multiple scenes or systems.