Do gaze disruptions indicate the perceived quality of non-uniformly coded natural scenes?

Abstract

Subjective video quality evaluation techniques usually involves a subject voluntarily attending to several regions in a video in order to scrutinize its quality. These techniques often tend to over-estimate the visual thresholds and in cases of non-uniform quality/coding, subjects mostly perceive the underlying video content in worse quality. This occurs due to a process known as Attentional Modulation occurring in the higher visual cortical areas V2 to V4. Examining disruptions in free viewing gaze patterns on the other hand, are said to be a more naturalistic method to measure the perceived video quality in such cases. To explore the feasibility of such a gaze disruption based quality metric, we examine the dependency between the two indicators: gaze disruptions and perceived subjective quality, obtained from a carefully controlled subjective test. By the examination of eye-tracking data and subsequent statistical analysis of difference opinion scores given by users, we are able to see that disruptions are indeed excellent indicators of perceived quality achieving a correlation 0.84. Several state of the art objective video quality metrics like SSIM, VIFp, VQM and PSNR-HVS (designed mostly for evaluation of uniform-quality) on the other hand, only produce a correlation ranging from 0.01 to 0.10. We conclude therefore that gaze disruptions may be used as excellent natural indicators of perceived quality in cases where quality is non-uniform, and may serve as new ground truth indicators for objective algorithms like Scanpath disruption(prediction) metrics that measure video quality in a more real-world like(naturalistic) manner.