Comparison of Generative Neural Networks as Models of Eye Movements

Saliency in psychology and computer vision is a measure of how much any particular point of an image attracts human attention. It can be measured with eye movements using eye-trackers: first, visual fixations of a human on particular points are saved, and later they are aggregated into a single gray-scale image and smoothed by a Gaussian kernel. There are many models that generate saliency maps while only taking an image as an input. There are 83 such models currently listed on MIT saliency benchmark. They range from biologically plausible to ones that focus on high accuracy of prediction. It is possible to define the problem of generating saliency maps as a problem of image-to-image translation between images representing scenes in real world and images representing saliency maps corresponding to these scenes. This problem has general solutions and it is a field of active research in computer vision. As a framework for solving this general task a model called "pix2pix" was proposed, which is used as a baseline in similar and derivative tasks such as photographic image synthesis. This work evaluates performance of pix2pix as a saliency map generation model and gives a comparison of different generative models based on neural networks at this task.