Did You Notice It?—How Can We Predict the Subjective Detection of Video Quality Changes From Eye Movements?

In the case of videos, the measurement of subjective detection of quality degradation usually needs observers to report the detection. However, if this detection could be automatically deduced by nonintrusive means, viewing would not be disturbed. This study examined how the subjective detection of quality degradation can be seen from eye movements, both in terms of group averages as well as from cases of single viewing times. In the experiment, the subjects were given the task of detecting local quality degradation in a video and reporting its significance to the quality of a video. The results of group averages showed that detection of the degraded area changed the viewing strategy from a search with short fixation to an evaluation with long fixations; most importantly, once the evaluation phase had begun, fixations concentrated on a smaller area than in the search phase. The change in eye movements was due to detection and not the quality level. The change in task was also seen in the eye movement measurements from single viewings, the spatial entropy of fixations being the best predictor of subjective detection, with a prediction accuracy rate of 78%. The combination of a temporal change in saccade amplitude and fixation duration averages also predicted detection in single viewing cases, with an accuracy of 68% The study discusses the use of eye movement measurements in video quality estimation and demonstrates that it is possible to predict subjective perception based on the stage of the task.