PURPOSE

Humans have evolved an efficient eye movement system for sampling the visual world. Last year we introduced a shape learning task and modeling approach for probing the algorithms that underlie eye movements (Renninger, Verghese & Coughlan, 2005). In that work, we discerned that despite the unfamiliarity of the shapes used in the task, observers spontaneously perceived each shape as having multiple "parts". To what extent do part representations play a role in eye movement planning?

METHODS

Observers actively studied novel shapes as their eye movements were measured. After this learning interval, the shape was displayed with a highly similar shape and the observer selected which one they studied. We compare eye movement traces with prediction maps from several theoretical models, using a rigorous SDT approach. We include a model that makes fixations to the informative “parts” of the shape. In a separate experiment, perceived parts were marked by observers using a manual segmentation tool. Information is defined as the uncertainty (entropy) of edge orientation distributions along the bounding contour of the shape or shape part.

RESULTS & CONCLUSIONS

Despite their good performance in the discrimination task, observers were not optimal when planning eye movements to the shape. Saliency was also a poor strategy for describing where fixations will occur. Strategies that look at "informative parts" or at locations that "reduce local orientation uncertainty" are the best predictors of fixation locations in this task. We conclude that observers are using a hybrid of strategies when planning eye movements.