Creating saliency maps for video stimuli and comparing with eye movement data

Abstract

The paper presents a deep-learning model that may be used to calculate saliency maps for video content. Classic saliency algorithms take into account only spatial information obtained from an image to generate a saliency map showing regions of higher importance—hence, regions where people are more likely to turn their gaze and attention. Algorithms for video stimuli add temporal information about the movements of objects on a frame-to-frame basis, resulting in more complex three-dimensional architectures. The paper analyses existing models and proposes a model based on one of them. The model’s performance is compared with the literature using four widely accepted measures (AUC, NSS, SIM, and CC). It is comparable and, in many cases, even better than already published models. Additionally, because of some improvements in the architecture, it is significantly faster in terms of the number of frames processed per second.