Attention Tracking using Shared Attention Modeling and Attentional Push

点击翻译按钮获取中文摘要

Our research program is concerned with attention tracking - determining where, and to what, people are paying attention. In particular, we are interested in measuring and predicting the visual attention of people viewing static photographs, computer displays, but also while watching videos or cinematic movies. Existing attention tracking methods based on eye tracking work well, but can be improved. One area of improvement was examined in recent work by our research group which concerned the inference of viewer task and its relation to the allocation of attention. Knowing what task a viewer is doing, such as reading, counting, or searching, is crucial to fine-tuning the notion of what is salient for that viewer. However, even this is not enough - a single randomly selected human will still do a better job of predicting where another human will look than even the best current attention tracking algorithms. One of the shortcomings of the current approaches is that, for the most part, they concentrate on analyzing regions of the image for their power to attract attention. The research program being proposed goes beyond these approaches and instead of only computing the power of an image region to 'pull' attention to it, we also consider the strength with which other regions of the image 'push' attention to the region in question. Objects or features in the image are analyzed as active manipulators of the viewers' attention. For example, man-made arrows on signs purposefully direct the viewer where to look, while converging lines in photographs can draw the viewers' attention to the intended subject in the image. Faces in images are known to be particularly effective in directing viewers attention in the direction of the face's gaze. We use the term 'Attentional Push' to refer to the power of image regions to direct and manipulate the attention allocation of the viewer. Our research will develop techniques for computing Attentional Push, which can then be integrated with standard image salience-based attention modeling algorithms to improve the ability to predict where viewers will fixate. In developing our techniques for computing Attentional Push, we will use knowledge gained from studies of 'Shared Attention' in humans and robots. Shared Attention is the process by which multiple 'actors' come to focus their attention on the same things. Our approach to Shared Attention is to identify the actors in the image, which can then be analyzed for their Attentional Push, potentially directing and manipulating the attention allocation of the viewer. The proposed work has many practical applications beyond understanding the attention of image viewers. We can use the methods developed to determine the collective focus of shared attention in surveillance of areas where many people gather, which could be used to analyze videos of group activities, such as sporting events, and predict the attention of viewers of these videos.