Abstract
We have investigated this question for static 2D images of natural scenes,and for affine transformations, specifically translation, rotation, scaling and shear. Affine transformations preserve lines and parallelism within an image regardless of the level of transformation. Some of these transformations, such as translation and rotation, are a natural part of our visual experience, whereas others, such as shear, are not. Human test subjects were required to discriminate pairs of natural scenes in which one image was a transformed version of the other. In order to compare the subjects responses to the different transformations we used two common measures: one based on the Euclidean distance of the pixel intensity values and the other based on the Euclidean distance of the pixel positions. Regardless of the metric used, for these affine transformations, the results suggest that humans are most sensitive to rotation. However, not all transformations (e.g., additive noise) are well described by a position-distance metric, therefore we describe results that compare affine transformations with other types of transformations. We discuss results in relation to the probability that such a transformation will occur in a natural scene. These results have allowed us to formulate a general model that predicts the sensitivity to different kind of transformations in images of natural scenes.
Research supported by the Canadian Institute of Health. Research grant MOP-11554 given to Fred Kingdom.