Abstract
In the real world, we face a constant flow of visual features. Our visual system has to cope with it, but its capacity is too limited to grasp all the visual features at a glance. Fortunately, information in the real world is often redundant. Aggregating inherent property of visual features as its statistical summary would therefore be efficient in representing the complex real world. Many studies suggest that we are able to extract some statistical summary, such as mean, variance, etc. from a single visual feature under brief viewing. It is also likely that we could extract covariation, a statistical description on multiple features, which would also greatly help to form a compact representation of the visual world. We examined the human sensitivity to the statistical covariation embedded in a visual scene display consisting of several circles filled with a sinusoidal grating. Different circles had different sizes and sinusoidal orientations, which assigned three types of covariation between features: size-orientation, orientation-location, and size-location. The degree of covariation was manipulated by varying the Pearson correlation coefficient. Participants sequentially viewed two 250-ms displays, and then judged which display contained the more correlated features. One display had a perfect correlation (r = 1.0) and the other not (r = 0, 0.33, or 0.67). Our results showed that the correct response performance was significantly lower for the size-orientation combination (slightly above chance level) than those for the other two location-related combinations. This was robust for various experimental conditions, namely varying ranges of stimulus values, reversing the direction of correlation, and reducing the top-down task set: All showed robustness for the lower performance for size-orientation and higher one for location-related correlation judgments. This suggests that, under brief viewing, human perception of statistical covariation among multiple features is founded on location dimension.
Meeting abstract presented at VSS 2016