視線運動解析による興味アスペクトの推定

抄録

This paper introduces a novel model for estimating the aspects of items that we are interested in (aspects of interest) when browsing contents. The proposed model generates items of gaze from the aspects of interest, where the aspects are characterized by associating them with attributes of items such as their specifications and appearances, and the aspects of interest are modeled by the mixtures of the aspects. Then, we can achieve aspects in a data-driven manner and represent aspects of interest flexibly via the estimation of the proposed model. Once we learn the aspects, we can also estimate the aspects of interest from newly observed gaze data. Furthermore, we incorporate influences of content layouts upon regions of gaze such as high conspicuity of center regions. Our experiments demonstrate the appropriateness of proposed model by accurately predicting items that users are interested in and likely to be looked at from estimated aspects of interest.