Information-theoretic model comparison unifies saliency metrics

Significance Where do people look in images? Predicting eye movements from images is an active field of study, with more than 50 quantitative prediction models competing to explain scene viewing behavior. Yet the rules for this competition are unclear. Using a principled metric for model comparison (information gain), we quantify progress in the field and show how formulating the models probabilistically resolves discrepancies in other metrics. We have also developed model assessment tools to reveal where models fail on the database, image, and pixel levels. These tools will facilitate future advances in saliency modeling and are made freely available in an open source software framework (www.bethgelab.org/code/pysaliency). Learning the properties of an image associated with human gaze placement is important both for understanding how biological systems explore the environment and for computer vision applications. There is a large literature on quantitative eye movement models that seeks to predict fixations from images (sometimes termed “saliency” prediction). A major problem known to the field is that existing model comparison metrics give inconsistent results, causing confusion. We argue that the primary reason for these inconsistencies is because different metrics and models use different definitions of what a “saliency map” entails. For example, some metrics expect a model to account for image-independent central fixation bias whereas others will penalize a model that does. Here we bring saliency evaluation into the domain of information by framing fixation prediction models probabilistically and calculating information gain. We jointly optimize the scale, the center bias, and spatial blurring of all models within this framework. Evaluating existing metrics on these rephrased models produces almost perfect agreement in model rankings across the metrics. Model performance is separated from center bias and spatial blurring, avoiding the confounding of these factors in model comparison. We additionally provide a method to show where and how models fail to capture information in the fixations on the pixel level. These methods are readily extended to spatiotemporal models of fixation scanpaths, and we provide a software package to facilitate their use.