Seminar Series

Psychophysics of natural images

The paradox of vision science is that it appears easy to learn much about very simple image viewing situations, or else little about the perception of more complex scenes. Since we are interested in the perception of natural images, we are adapting normal psychophysical methods to study such scenes. We show that there is an optimisation to the properties of natural scene statistics. In addition, we find that simple, low-level models can successfully predict psychophysical results using naturalistic stimuli. This implies that such low-level processes can provide useful insight into the performance a wide variety of tasks. We are currently extending this approach to deal with more complex, dynamic, stimulus sequences in which the task is to predict whether violent human behaviour will occur. It may be that the primitives for such tasks can also be defined as simple, low-level operators. There may therefore be a "middle way" through the paradox: make the images complex, but keep the models simple. The real question, then, is defining the limits of this approach.