Many methods of statistical shape description operate by describing shapes in terms of the variations inherent in a training set. This represents a limitation in that a training set must be assembled beforehand, and that only shapes lying within the span of the training data can be succinctly described. We develop a statistical representation that describes a shape in terms of the variations inherent in that shape, without reference to training images. Our new representation is then used to characterise a number of perceptual deformations, with the intent being to investigate how well such deformations can be captured and modelled by our description.