Experimental metaphysics: The scanpath as an epistemological mechanism

State-of-the-art in visual attention modeling

Modeling visual attention--particularly stimulus-driven, saliency-based attention--has been a very active research area over the past 25 years. Many different models of attention are now available which, aside from lending theoretical contributions to other fields, have demonstrated successful applications in computer vision, mobile robotics, and cognitive systems. Here we review, from a computational perspective, the basic concepts of attention implemented in these models. We present a taxonomy of nearly 65 models, which provides a critical comparison of approaches, their capabilities, and shortcomings. In particular, 13 criteria derived from behavioral and computational studies are formulated for qualitative comparison of attention models. Furthermore, we address several challenging issues with models, including biological plausibility of the computations, correlation with eye movement datasets, bottom-up and top-down dissociation, and constructing meaningful performance measures. Finally, we highlight current research trends in attention modeling and provide insights for future.

Positional Features and Algorithmic Predictability of Visual Regions-of-Interest in Robot Hand Movement

Visual functions are important for robots who engage in cooperative work with other robots. In order to develop an effective visual function for robots, we investigate human visual scanpath features in a scene of robot hand movement. Human regions-of-interest (hROIs) are measured in psychophysical experiments and compared using a positional similarity index,Sp, on the basis of scanpath theory. Results show consistent hROI loci due to dominant top-down active looking in such a scene. This paper also discusses how bottom-up image processing algorithms (IPAs) are able to predict hROIs. We compare algorithmic regions-of-interest (aROIs) generated by IPAs, with the hROIs obtained from robot hand movement images. Results suggest that bottom-up IPAs with support size almost equal to fovea size have a high ability to predict the hROIs.

A model of modes of attention and inattention for artificial perception

This paper considers several aspects of natural visual attention and its link to wider notions of awareness, natural and artificial, in the context of foveated vision. It builds on a theory of abductive perception; a formal definition for an artificial or robot perceptual system, using objects represented as feature clouds. It proposes a broad, but unifying approach to several aspects of visual attention in the light of this, including autonomic eye gaze movements, aspects of secondary and covert attention, and exogenous (sense driven) and endogenous (task driven) attention. Modes of attentional lapse, commonly referred to as inattentional blindness and change blindness, are also discussed in the context of the model presented.

On the discriminability of hROIs, human visually selected regions-of-interest

Many studies have tried to answer an important question: is it possible to predict human visually selected regions-of-interest (hROIs)? hROIs are defined as the loci of eye fixations and they can be analyzed by their spatial distribution over the visual stimulus and their temporal ordering. We used a simplified set of geometrical spatial kernels and linear filter models as bottom-up conspicuity operators that produce algorithmically selected regions-of-interest, aROIs. As a direct approach we measured the ability of these aROIs to predict human scanpaths. The level of prediction is measured by two similarity indices: Sp for spatial similarity and Ss for temporal ordering similarity. At the same time we assessed the discriminability of the hROI loci, in terms of conspicuity, with respect to non-selected (not of interest) regions of an image. We prove that this discrimination is possible and further correlates with the positional similarity index Sp. Other human scanpath experimental conditions are presented in parsing diagrams and discussed. A general top-down/bottom-up scanpath model is finally formulated.

Oculomotor strategies for the direction of gaze tested with a real-world activity

Laboratory-based models of oculomotor strategy that differ in the amount and type of top-down information were evaluated against a baseline case of random scanning for predicting the gaze patterns of subjects performing a real-world activity--walking to a target. Images of four subjects' eyes and field of view were simultaneously recorded as they performed the mobility task. Offline analyses generated movies of the eye on scene and a categorization scheme was used to classify the locations of the fixations. Frames from each subject's eye-on-scene movie served as input to the models, and the location of each model's predicted fixations was classified using the same categorization scheme. The results showed that models with no top-down information (visual salience model) or with only coarse feature information performed no better than a random scanner; the models' ordered fixation locations (gaze pattern) matched less than a quarter of the subjects' gaze patterns. A model that used only geographic information outperformed the random scanner and matched approximately a third of the gaze patterns. The best performance was obtained from an oculomotor strategy that used both coarse feature and geographic information, matching nearly half the gaze patterns (48%). Thus, a model that uses top-down information about a target's coarse features and general vicinity does a fairly good job predicting fixation behavior, but it does not fully specify the gaze pattern of a subject walking to a target. Additional information is required, perhaps in the form of finer feature information or knowledge of a task's procedure.

Human-vision-based selection of image processing algorithms for planetary exploration

Independent and automatic image processing is a fundamental objective of the computer vision community. Understanding the role of the eye movement scan path in human vision is an important step toward the achievement of this objective. This top-down model of higher human vision is a new approach to bottom-up image processing algorithms and provides an important new metric and tool in computer vision. We have demonstrated that a small and manageable collection of image processing algorithms, experimentally selected and then combined together can serve in a task such as predicting human eye fixations identifying geological features. Thus, automatic picture analysis based upon human vision could be an essential element in planetary exploration.

Fixation behavior while walking: persons with central visual field loss

The aim of this study was to determine the effect of central visual field loss (CFL) on fixation patterns of a person walking towards a target. Subjects were four visually normal persons and 10 persons with CFL. Eye position on scene was recorded and classified into 20 scene categories. The distributions of fixations among scene categories were compared across the two subject groups. For all but two CFL subjects, who fixated primarily at the floor, the distributions of fixations for the CFL subjects ranged from being moderately to strongly correlated with that of the visually normal mean. An analysis of the similarity in the sequence of fixations (or gaze pattern) of the CFL subjects to the visually normal subjects showed a range of 7-66%. Excluding the one CFL subject who had a functioning fovea, sequence similarity was strongly correlated with the logarithm of the minimum angle of resolution (logMAR). The better a person's logMAR, the more closely his or her gaze pattern matched that of the visually normal subjects. Finally, the CFL data were tested against two current models of oculomotor strategy, visual salience and guided search. Similar to what was found with visually normal subjects, CFL subjects appear to use the expected features and general location of the target to guide their fixations, the guided-search strategy.

Computational modelling of visual attention

Five important trends have emerged from recent work on computational models of focal visual attention that emphasize the bottom-up, image-based control of attentional deployment. First, the perceptual saliency of stimuli critically depends on the surrounding context. Second, a unique 'saliency map' that topographically encodes for stimulus conspicuity over the visual scene has proved to be an efficient and plausible bottom-up control strategy. Third, inhibition of return, the process by which the currently attended location is prevented from being attended again, is a crucial element of attentional deployment. Fourth, attention and eye movements tightly interplay, posing computational challenges with respect to the coordinate system used to control attention. And last, scene understanding and object recognition strongly constrain the selection of attended locations. Insights from these five key areas provide a framework for a computational and neurobiological understanding of visual attention.