How similar are the eye movement patterns of different subjects when free viewing dynamic natural scenes? We collected a large database of eye movements from 54 subjects on 18 high-resolution videos of outdoor scenes and measured their variability using the Normalized Scanpath Saliency, which we extended to the temporal domain. Even though up to about 80% of subjects looked at the same image region in some video parts, variability usually was much greater. Eye movements on natural movies were then compared with eye movements in several control conditions. "Stop-motion" movies had almost identical semantic content as the original videos but lacked continuous motion. Hollywood action movie trailers were used to probe the upper limit of eye movement coherence that can be achieved by deliberate camera work, scene cuts, etc. In a "repetitive" condition, subjects viewed the same movies ten times each over the course of 2 days. Results show several systematic differences between conditions both for general eye movement parameters such as saccade amplitude and fixation duration and for eye movement variability. Most importantly, eye movements on static images are initially driven by stimulus onset effects and later, more so than on continuous videos, by subject-specific idiosyncrasies; eye movements on Hollywood movies are significantly more coherent than those on natural movies. We conclude that the stimuli types often used in laboratory experiments, static images and professionally cut material, are not very representative of natural viewing behavior. All stimuli and gaze data are publicly available at http://www.inb.uni-luebeck.de/tools-demos/gaze.

Saliency prediction typically relies on hand-crafted (multiscale) features that are combined in different ways to form a "master" saliency map, which encodes local image conspicuity. Recent improvements to the state of the art on standard benchmarks such as MIT1003 have been achieved mostly by incrementally adding more and more hand-tuned features (such as car or face detectors) to existing models [18,4,22,34]. In contrast, we here follow an entirely automatic data-driven approach that performs a large-scale search for optimal features. We identify those instances of a richly-parameterized bio-inspired model family (hierarchical neuromorphic networks) that successfully predict image saliency. Because of the high dimensionality of this parameter space, we use automated hyperparameter optimization to efficiently guide the search. The optimal blend of such multilayer features combined with a simple linear classifier achieves excellent performance on several image saliency benchmarks. Our models outperform the state of the art on MIT1003, on which features and classifiers are learned. Without additional training, these models generalize well to two other image saliency data sets, Toronto and NUSEF, despite their different image content. Finally, our algorithm scores best of all the 23 models evaluated to date on the MIT300 saliency challenge [16], which uses a hidden test set to facilitate an unbiased comparison.

No abstract