Best-first interpretation of imperfect object contours
Bernd Neumann · 1978
A scene-analysis system is presented which identifies objects by means of straight-line contour approximations. Descriptions of prototypes are stored in a relational-model database. Scenes may be disturbed and may contain objects which are partially occluded by another object. A segmentation algorithm extracts straight edges by multidirectional sampling of the picture gradient matrix. Interpretation hypotheses are generated using 2-step relational matching. A best-first strategy selects the final interpretation using constraint filtering for ambiguous situations. Results are presented for scenes containing configurations of artificial shapes (card-board pieces) and real-life objects (tools). They indicate that in systems designed for application one should sophisticate interpretation rather than segmentation.