Determining comprehension and quality of TV programs using eye-gaze tracking

Currently, TV programs are evaluated by using questionnaires given after previews or by using TV ratings. There are few objective criteria useful for describing technical know-how about program production. One of the TV program producers' concerns is how to choose expression methods that convey their ideas to viewers correctly and efficiently. Research has shown that eye-gaze direction is related to the human focus and attention. Gaze-based evaluations have been proposed for image-quality evaluations and certain usability tests. Such approaches are mainly based on how often a specific region attracted the subjects' gaze or how long their gaze was fixed on it. To apply these approaches to TV programs, all the object regions that seem to attract a viewer's gaze need to be specified in advance. This causes several problems including the accuracy of specifying the region by using an image processing technique is not equal to the human subject's recognition ability and it is not feasible to manually specify such regions in an enormous number of frames (images) comprising the program. Further, how characteristics of well-produced TV programs appear on the viewer's gaze has not been objectively analyzed yet. There is a need to investigate the relationship between gaze and program contents which can be used as means for improving comprehension and quality of the TV programs. In this paper, we propose a new measurement and evaluating method for this purpose. This paper focuses on the relationship between a viewer's comprehension of a program and their gaze direction in a real experimental TV educational program involving 26 elementary school children and broadcast by NHK Broadcasting Corporation of Japan. Correlation between TV program comprehension and entropy is investigated. That is, variances in the gaze direction in relation to program comprehension are based on a entropy value that represents the degree of dispersion in each frame and is calculated from a probability density function estimated from the gaze directions. The results indicate that the variances of the gaze direction for scenes that gave better comprehension tended to be lower. This tendency was further noticeable after a keyword utterance were related to the answers of corresponding questions. (c) 2007 Elsevier Ltd. All rights reserved.