Scanpaths Generation for Target Search Based on Deep Learning

Scanpath is a sequence of gaze fixations changing with time when browsing something, which records eyes' movement dynamically. Accurate prediction of scanpaths can help computers better predict which area human pay attention to and drive the development of next-generation systems that can understand human behavior and needs. Therefore, this paper introduces an algorithm of scanpaths generation for target search. Based on one-shot learning network, the algorithm extracts the target map with task information, and then imports it into a visual model of the superior colliculus to predict the task-drive scanpaths. The results are similar to the gaze behavior of human eyes in shape, direction and length. At the same time, the algorithm also has the ability to predict the scanpaths based on unseen categories.