Head-pose-based attention recognition on large public displays

Estimating the attention to regions of large public displays is a problem that has been present since their very advent. In contrast to traditional screen interaction with a web browser, no clickstream can be calculated for such displays. With means to estimate where the users look at, this issue could be partly overcome. In this work we evaluate di?erent aspects (head movement, individual persons, locations, amount of training data) and their impact on the recognition accuracy of the viewing direction based on head pose only. Statistic evaluation revealed, for example, that (i) free head movement in yaw and pitch directions has insignificant influence on the recognition accuracy as compared to limited movement in vertical/horizontal direction only, (ii) high differences of up to 16% in the accuracy suggest that such a system should be trained per person to achieve optimum recognition performances, and (iii) calibration on multiple positions does not enhance recognition rate significantly as compared to training on only a single spot.