Characterizing eye movements during temporal and global quality assessment of h.264 compressed video sequences

摘要

Studies have shown that the assessment of image or video quality is closely linked to the deployment of visual attention measured through the eye movements of observers. However, this link is not fully understood yet. While most studies have considered the evaluation of global quality (i.e. produced by all types of defects) this study focuses on the evaluation of what we call temporal quality (i.e. produced by defects variating from one frame to another). In this study, we measured whether assessing the temporal quality of a video influences the visual exploration of the corresponding video.We set-up a subjective experiment in which the eye movements of observers are recorded during three different tasks: a free-viewing task (FT), a global quality assessment task (QA-G) and a temporal quality assessment task (QA-T). The FT acts as a reference representing the standard visual exploration for our video database in perfect quality. Eye movements recorded during this task are compared to those obtained during the two quality assessment tasks for each sequence at different quality levels.As previously shown, we found that observers assessing global quality gaze at locations dissimilar to those fixated during the FT. It is however not the case for locations fixated by observers assessing temporal quality. The fixated locations from QA-T are more similar to the ones from FT. Morever, our results suggest that video quality level does not influence the locations fixated but has an impact on the fixation durations: the lower the quality is, the longer observers maintain their fixations. Finally, fixation locations are less variable between observers during the quality assessment tasks than during the FT for either perfect or poor quality level.We also analyzed the evolution over time of these eye movement parameters (fixation locations and durations). It appears that all indicators are similar for the three tasks during the first 1 or 2 seconds of the videos and that differences between tasks appear later on. Hence, it seems that the task to assess video quality plays a role afterwards on the deployment of the participants' visual attention.