Saliency and eye movements in the perception of natural scenes

Humans inspect the environment around them by selecting a sequence of locations to fixate which will provide information about the scene. How are these locations chosen? The saliency map model suggests that points in the scene are represented topographically and that the likelihood of them being fixated depends on low-level feature contrast. This model makes specific predictions about the way people will move their eyes when looking at natural scenes, although there are few experimental tests of these predictions. The experiments described in this thesis show effects of visual saliency on the likelihood and the speed at which objects are fixated. Experiment 1 shows that the potency of salient objects is moderated by the task being performed. When the task does not constrain the regions of interest, as in a general encoding situation, the saliency model performs better than chance estimates (Experiments 2 and 3). There are also sequential patterns of eye movements in this task - scanpaths - that the model does not reproduce. In visual search, participants can saccade to a target object, and this is quicker, in some cases, if the target is more salient (Experiments 4-7). A salient distractor impedes search more than a non-salient one (Experiments 8 and 9). The context of the scene also has an effect on search, and features of the layout, in particular the horizon, may cause an asymmetry in saccade direction (Experiment 10). Findings from research with a visual agnosia patient are consistent with the idea that scene understanding and saliency combine in guiding the eyes (Experiment 11). These experiments support a framework that incorporates a task-driven prior, gist and the relevance of each region to the task, in addition to bottom-up saliency. Thus saliency is just one part of the way in which people move their eyes