Human scanpath estimation based on semantic segmentation guided by common eye fixation behaviors

摘要

To explore the dynamic process of complex human eye movement behavior, we proposed a new model to simulate human scanpath when subjects observed natural images freely. Previous methods almost focused on finding effective and advanced technology, such as machine learning or deep learning, for estimating human scanpath. In contrast, our proposed method devoted to find a new way that could use the intrinsic property of eye-tracking data between different races to guide the design of a deep network. Inspired by that, the model of human scanpath estimation was established, which based on a semantic segmentation module guided by common eye fixation behaviors between people with different cultures. The semantic segmentation module could deal with locating fixations positions and the fixations ranking problem in parallel and generate human scanpath combined with the output of common attention portions (CAP) generator. The common attention portions (CAP) generator was designed to optimize the performance of semantic segmentation module and extract the common eye fixation behaviors between people with different cultures. We evaluated the performance of our model on three public eye tracking datasets by comparing the result generated from our model with the ground truth of scanpath produced by a new method in this work. The proposed model also achieved the encouraging performance compared with some classic and fashionable models.(c) 2020 Elsevier B.V. All rights reserved.