Towards Computational Identification of Visual Attention on Interactive Tabletops

There is a growing interest in the ability to detect where people are looking in real-time to support learning, collaboration, and efficiency. Here we present an overview of computational methods for accurately classifying the area of visual attention on a horizontal surface that we use to represent an interactive display (i.e. tabletop). We propose a new model that utilizes a neural network to estimate the area of visual attention, and provide a close examination of the factors that contribute to the accuracy of the model. Additionally, we discuss the use of this technique to model joint visual attention in collaboration. We achieved a mean classification accuracy of 75.75% with a standard deviation of 0.14 when data from four participants was used in training the model and then tested on the fifth participant. We also achieved a mean classification accuracy of 98.8% with 0.02 standard deviation when different amounts of overall data was used to test the model.