This paper describes the POLL (Perceptual Organization and Line Labeling) system for obtaining labeled line drawings from single intensity images using an integrated blackboard system. This system emphasizes the data driven extraction of 3D geometric information from intensity images. The robustness of the system comes from it being implemented in an integrated framework in which the errors made by one module can be diagnosed and corrected by the constraints imposed by other modules. The system is able to generate an initial line drawing from an intensity image in the domain of piecewise smooth objects.
Four modules were used as knowledge sources: weak membrane edge detection, curvilinear grouping, proximity grouping, and curvilinear line labeling. An initial representation of the image data is built using the first three knowledge sources. This representation is analyzed using a modified curvilinear line labeling algorithm developed in this paper that uses figure-ground separation to constrain legal line labelings. This modified line labeling algorithm can diagnose problems with the initial representation. The modified line labeling algorithm can find errors such as missing edges, improperly typed vertices, and missing phantom junctions. Errors in the representation can be fixed using a set of heuristics that were created to repair common mistakes. If no irreparable errors are found in the representation, then the modified line labeling algorithm produces a 3D interpretation of the data in the input image without explicitly reconstructing 3D shape.