Predicting Visual Fixations.

As we navigate and behave in the world, we are constantly deciding, a few times per second, where to look next. The outcomes of these decisions in response to visual input are comparatively easy to measure as trajectories of eye movements, offering insight into many unconscious and conscious visual and cognitive processes. In this article, we review recent advances in predicting where we look. We focus on evaluating and comparing models: How can we consistently measure how well models predict eye movements, and how can we judge the contribution of different mechanisms? Probabilistic models facilitate a unified approach to fixation prediction that allows us to use explainable information explained to compare different models across different settings, such as static and video saliency, as well as scanpath prediction. We review how the large variety of saliency maps and scanpath models can be translated into this unifying framework, how much different factors contribute, and how we can select the most informative examples for model comparison. We conclude that the universal scale of information gain offers a powerful tool for the inspection of candidate mechanisms and experimental design that helps us understand the continual decision-making process that determines where we look.

DeepGaze III: Modeling free-viewing human scanpaths with deep learning.

Humans typically move their eyes in "scanpaths" of fixations linked by saccades. Here we present DeepGaze III, a new model that predicts the spatial location of consecutive fixations in a free-viewing scanpath over static images. DeepGaze III is a deep learning-based model that combines image information with information about the previous fixation history to predict where a participant might fixate next. As a high-capacity and flexible model, DeepGaze III captures many relevant patterns in the human scanpath data, setting a new state of the art in the MIT300 dataset and thereby providing insight into how much information in scanpaths across observers exists in the first place. We use this insight to assess the importance of mechanisms implemented in simpler, interpretable models for fixation selection. Due to its architecture, DeepGaze III allows us to disentangle several factors that play an important role in fixation selection, such as the interplay of scene content and scanpath history. The modular nature of DeepGaze III allows us to conduct ablation studies, which show that scene content has a stronger effect on fixation selection than previous scanpath history in our main dataset. In addition, we can use the model to identify scenes for which the relative importance of these sources of information differs most. These data-driven insights would be difficult to accomplish with simpler models that do not have the computational capacity to capture such patterns, demonstrating an example of how deep learning advances can be used to contribute to scientific understanding.

Semantic object-scene inconsistencies affect eye movements, but not in the way predicted by contextualized meaning maps.

Semantic information is important in eye movement control. An important semantic influence on gaze guidance relates to object-scene relationships: objects that are semantically inconsistent with the scene attract more fixations than consistent objects. One interpretation of this effect is that fixations are driven toward inconsistent objects because they are semantically more informative. We tested this explanation using contextualized meaning maps, a method that is based on crowd-sourced ratings to quantify the spatial distribution of context-sensitive "meaning" in images. In Experiment 1, we compared gaze data and contextualized meaning maps for images, in which objects-scene consistency was manipulated. Observers fixated more on inconsistent versus consistent objects. However, contextualized meaning maps did not assign higher meaning to image regions that contained semantic inconsistencies. In Experiment 2, a large number of raters evaluated image-regions, which were deliberately selected for their content and expected meaningfulness. The results suggest that the same scene locations were experienced as slightly less meaningful when they contained inconsistent compared to consistent objects. In summary, we demonstrated that - in the context of our rating task - semantically inconsistent objects are experienced as less meaningful than their consistent counterparts and that contextualized meaning maps do not capture prototypical influences of image meaning on gaze guidance.

There is no evidence that meaning maps capture semantic information relevant to gaze guidance: Reply to Henderson, Hayes, Peacock, and Rehrig (2021).

The concerns raised by Henderson, Hayes, Peacock, and Rehrig (2021) are based on misconceptions of our work. We show that Meaning Maps (MMs) do not predict gaze guidance better than a state-of-the-art saliency model that is based on semantically-neutral, high-level features. We argue that there is therefore no evidence to date that MMs index anything beyond these features. Furthermore, we show that although alterations in meaning cause changes in gaze guidance, MMs fail to capture these alterations. We agree that semantic information is important in the guidance of eye-movements, but the contribution of MMs for understanding its role remains elusive.

Meaning maps and saliency models based on deep convolutional neural networks are insensitive to image meaning when predicting human fixations.

Eye movements are vital for human vision, and it is therefore important to understand how observers decide where to look. Meaning maps (MMs), a technique to capture the distribution of semantic information across an image, have recently been proposed to support the hypothesis that meaning rather than image features guides human gaze. MMs have the potential to be an important tool far beyond eye-movements research. Here, we examine central assumptions underlying MMs. First, we compared the performance of MMs in predicting fixations to saliency models, showing that DeepGaze II - a deep neural network trained to predict fixations based on high-level features rather than meaning - outperforms MMs. Second, we show that whereas human observers respond to changes in meaning induced by manipulating object-context relationships, MMs and DeepGaze II do not. Together, these findings challenge central assumptions underlying the use of MMs to measure the distribution of meaning in images.

Information-theoretic model comparison unifies saliency metrics.

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.