Decoding What People See from Where They Look: Predicting Visual Stimuli from Scanpaths

摘要

Saliency algorithms are applied to correlate with the overt attentional shifts, corresponding to eye movements, made by observers viewing an image. In this study, we investigated if saliency maps could be used to predict which image observers were viewing given only scanpath data. The results were strong: in an experiment with 441 trials, each consisting of 2 images with scanpath data - pooled over 9 subjects - belonging to one unknown image in the set, in 304 trials (69%) the correct image was selected, a fraction significantly above chance, but much lower than the correctness rate achieved using scanpaths from individual subjects, which was 82.4%. This leads us to propose a new metric for quantifying the importance of saliency map features, based on discriminability between images, as well as a new method for comparing present saliency map efficacy metrics. This has potential application for other kinds of predictions, e.g., categories of image content, or even subject class.