Morphological shape analysis of children's actively generated viewpoints

Abstract

Our visual system develops in a world of three-dimensional objects, where children often control the dynamic visual experience they experience. Our previous experimental work was the first to analyse in detail what object views are selected in children’s unconstrained manipulation of physical objects. The key finding was the distribution of dwell time across all possible viewpoints; preferred viewpoints are around on-axis views, where the principal axis of elongation is either perpendicular or parallel to the line of sight and flat surfaces are presented approximately perpendicular to the line of sight – so called planar views. The planar bias is statistically above a random exploration baseline at the earliest age tested (12-18 months) and undergoes large changes until at least 30 to 36 months. The planar bias is also characteristic of mature viewing and was found to promote more efficient visual object learning. Its functional role is, however, unknown. Here we present a morphological shape analysis of the object views selected by a large cross-sectional sample of children, while they manipulated objects (N = 54, 12-36 months). We followed an image-based (pixel-based) approach; we coded the 3D object orientation and used it to create computer-generated images of the object and extracted standard shape measures. We investigated if basic shape measures differed between planar and non-planar views. In particular we computed surface area and aspect ratio. In addition, and for a smaller subset of 30-36 month olds (N = 8, 14400 frames) we computed the number of pixels that do not overlap between consecutive frames (a measure of image variability), as a proportion of the maximum object surface area. There were two key results: (1) on average, planar views have lower surface area and lower aspect ratio (i.e. they are more elongated) than non-planar views; (2) on average, consecutive frames of planar views have a lower proportion number of pixels that do not overlap between frames than non planar views. The results suggest that elongation is a characteristic aspect of preferred viewpoints, not solely of the 3D orientation of the main axis but also of the viewed surface. Also, planar views correspond to more stable periods of the object manipulation. This suggests that sampling planar views corresponds to moments of focused attention to a particular view where perhaps learning of a static view is occurring. We will discuss these results in the context of contemporary theories of visual object recognition development.