Fine-tuning SalGAN and PathGAN for extending saliency map and gaze path prediction from natural images to websites

Abstract

In recent years, researches dealing with the study of visual attention have become very popular thanks to the enormous increase of Artificial Intelligence. Machine Learning and, in particular, Deep Learning allowed researchers to propose new predictive models operating on natural images. In the meantime, an increasing number of websites has been made available on the Internet. However, few approaches, aiming at extending the results obtained on natural images to web pages, have been proposed. In this paper, we provide a contribution in this setting by applying fine-tuning and other refinements to two existing GAN-based approaches (i.e., SalGAN and PathGAN) originally proposed to predict the saliency maps and gaze paths on natural images. Our ultimate goal is defining some variants of them able to deal with websites. In particular, our SalGAN variant represents one of the first attempts to employ GANs for saliency map prediction on web pages, whereas our PathGAN variant is the first attempt to adopt GANs for gaze path prediction on websites Here, we present our proposals, highlight their main novelties, describe the tests done and the results obtained. We also highlight two further contributions of this paper, namely: (i) a new dataset, more complete than the existing ones, supporting the analysis of visual attention on websites, and (ii) a tool supporting a web page designer in her attempt to increase the visitor interest and curiosity.