Human subjects can locate a target very quickly if the target is sufficiently different from distractors. This is visual pop-out.
Pop-out is only one of the many strategies used to direct visual attention.
Vision systems use attention to fixate a fovea on different parts of the scene till the target is found.
I am developing a Bayesian theory of attention and visual pop-out. The figure on the right shows a simulation of the theory. In fifty trials, the simulation was able to find the target in four fixations (on the average) even when the scene contained fifty distractors.
Most theories of attention are implementational -- that is, they explain how a neural net might implement the observed properties of visual attention. They do not explain why that particular attention strategy is useful.
The Bayesian attention theory is computational. Using probabilistic arguments it clearly shows why pop-out like behavior is important.
Just as in human vision, the Bayesian attention mechanism exhibits pop-out. When the distractors are significantly different from the object, the object is found in constant time.
When the distractors look similar to the object, the number of fixations required to find the object grows linearly with the number of objects.
All of this has been experimentally verified.
At the moment, we are writing it up.
The Bayesian attention mechanism is looking for the following object (a cardboard goldfish):
The attention mechanism finds the fish by a sequence by fixations.
Here's one sequence of fixations. Note that the fish is heavily occluded.
The fixations and the tail of the fish are outlined in the image.
Request references by email.
Hemant's Home Page