Human and primate retinae have a small high resolution fovea while the remaining periphery is sampled at very low resolution, so the key first step in creating a visual system that can provide meaningful information to the robot is directing its eye to appropriate objects in the world.
Building the visual system
Our paper Vincent, Baddeley, Troscianko & Gilchrist (2005) provided a succinct explanation of why the early visual system, retina through to primary visual cortex, is organised the way how it is. The objective of maximising the amount of visual information transmitted through the visual system whilst minimising the metabolic costs of this transmission is sufficient to explain many aspects of the early visual system, more so that any previous model.
Learning about human eye movements
In trying to build an eye movement system, it is important to understand the behaviour of the human eye. In Tatler & Vincent (submitted) we found that humans use a remarkable similar strategy of exploring visual scenes regardless of what particular the scene is actually of. This robust visual scanning strategy is implemented into the robotic eye.
Using vision in the fovea vs. periphery to drive saccades
One of the most obvious factors that may help guide the eye to appropriate objects in the world, is the visual information received. Many previous papers have shown that a variety of high spatial frequency feature map types, such as edges and colour, contribute to eye guidance. In Tatler, Baddeley & Vincent (2006), we showed that in fact high spatial frequency cues can only be used to guide relatively short eye movements; long-range eye movements are less guided by visual features. So the fovea/periphery arrangement of the retina really does need to be taken into account.
Assessing models of visually driven eye movements
We examined the dominant model of how eye movements are guided by visual features, the weighted salience model. By evaluating this model with a space-variant retina, we found that the models ability to direct the eye to target objects was particularly low Vincent, Troscianko & Gilchrist (submitted) and provided a poor fit to psychophysical data.
A Bayesian model of eye movements
To replace this, we propose an alternative view of eye movements but taking into account not only the space-variant retina, but also the role of top-down (or task-related) factors. This model is formulated in a Bayesian manner which is the best way to make inferences about target objects when the visual (and higher-level) knowledge is noisy and uncertain Vincent, Gilchrist, Troscianko (in progress).
Evaluating the robots eye movements
In order to evaluate how good the robotic eye movement system is, I developed a quantitative tool to evaluate various low- and high-level hypotheses about eye movements Vincent, Correani, Baddeley, Trocianko, Leonards (submitted). This allow us to empirically measure how well the robot visual system is doing in pointing the eye to appropriate areas of the visual scene in order to solve behavioural tasks.
The website is maintained by Richard Newcombe.