Towards Context-Dependence Eye Movements Prediction in Smart Meeting Rooms


Being able to predict gaze locations, as compared to only measuring them, is desirable in many systems such as the design of web pages and commercials adaptive user interfaces, interactive visualization, or attention management systems. However, accurately predicting eye movements remains a challenging problem. In this paper, we present the results of experimental study to improve the prediction of saliency maps in smart meeting rooms. More specifically, we investigate meeting scenarios in terms of their context-dependence saliency based on different image features. We have recorded the center of gaze of users in meeting rooms in different scenarios (giving a talk, listening). We then used a data-driven approach to find out which features are important in each scenario. We found that the predictions differ according to the type of features we selected. Most interestingly, we found that models trained on face features perform better than the models trained on other features in the giving a talk scenario, but in the listening scenario the models trained on competing saliency features from Itti and Koch perform better than the models trained on another features. This finding points towards including context information about the scene and situation into the computation of saliency maps as important towards developing models of eye movements, which operate well under natural conditions such as those encountered in ubiquitous computing settings.

Artificial Neural Networks and Machine Learning – ICANN 2014