Can Machines Interpret Line Drawings?

Abstract

Engineering design would be easier if a computer could interpret initial concept drawings. We outline an approach for automated interpretation of line drawings of polyhedra, and summarise what is already possible, what developments can be expected in the near future, and which areas remain problematic. We illustrate this with particular reference to our own system, RIBALD, summarising the published state of the art, and discussing recent unpublished improvements to RIBALD. In general, successful interpretation depends on two factors: the number of lines, and whether or not the drawing can be classified as a member of special shape class (e.g. an extrusion or normalon). The state-of-the-art achieves correct interpretation of extrusions of any size and most normalons of 20 30 lines, but drawings of only 10 20 lines can be problematic for unclassified objects.Despite successes, there are caseswhere the desired interpretation is obvious to a human but cannot be determined by currently-available algorithms. We give examples both of our successes and of typical caseswhere human skill cannot be replicated.