Identifying scanpath starting point in structured web images at group level Comparing mouse and eye tracking with saliency map

Abstract

Understanding where and when people look on webpages is essential to web creators. However, collecting gaze data with traditional eye tracking (ET) is expensive and time-consuming. Alpha.One, a neural marketing company, aims to predict the gaze sequence of viewing on webpages, using deep learning and generative adversarial neural networks (GANs). The models are trained on salience data which is aggregated from mouse tracking (MT) experiments on Amazon's Mechanical Turk. The experiments are conducted via a psychophysical paradigm known as the mouse-contingent multi-resolutional paradigm (Jiang et al., 2015). The hypothesis of this study is that the shifts of viewing order are initiated toward the salient intensity level (Henderson, 2003; Itti, 2005; Tseng & Howes, 2008; Underwood, 2009). This research presents a novel approach to (a) determine the starting point of where users are most likely to look at first on a webpage and (b) produce a general scanpath. The ET heat maps are compared to the starting point in general viewing order generated from ET and MT data. The results show the starting point usually is not in the most salient area of the ET heat maps, and the hypothesis that the first element to be looked at is in the most salient area is disproved. This indicates that the viewing order cannot be simply deduced from the salient intensity levels of the heat map