LEAD

Gary COTTRELL

Chercheur Invité LEAD CNRS UMR 5022
Université de Californie San Diego (UCSD), USA


SUN: A Model of Visual Salience Using Natural Statistics


As a result of having a foveated retina, we actively move our eyes in order to direct our highest resolution of visual processing towards interesting things. In fact, we move our eyes about three times a second; it is a decision we make about 172,000 times a day. How do we decide where to look?

Most computational theories of overt visual attention assume there is a salience map that is computed from both exogenous and endogenous sources. Exogenous sources of salience tend to be visually “busy” locations in the world. Endogenous sources of salience derive from task-relevant considerations, such as those involved in my most vexing task of the day, finding my glasses. We have developed a Bayesian model of salience that naturally leads to components corresponding to these two influences. Our model differs from most other models in assuming that for exogenous influences on salience, the statistics of the visual world are learned through experience and then applied to new stimuli, as opposed to being computed directly from new stimuli. The model has been applied to standard datasets and can account for eye fixations as well or better than any other model in the literature, is able to be computed in real time due to its computational efficiency, and is able to explain several visual search asymmetries.