Sketching: Estimating object positions from relational descriptions

Abstract

One of the long-standing research problems in computer vision is the conversion of spatial information between the concrete domain of position in an image or in coordinate space and the more abstract domain of symbolic scene descriptions. This paper discusses a technique for solving the layout problem, one side of this general conversion problem. From a scene description composed of a set of symbolic spatial relations with fuzzy truth values, estimates of the objects' positions are derived. The position estimates consist of coordinate intervals with fuzzy truth values expressing the confidence associated with each interval.