Segmentation and description of binocularly viewed contours

Contour segmentation is typically seen as the identification either of discondnuities'"? or of describable subsectionsIO operations which are usually treated as dual. As we are interested in providing primitive descriptions to higher processes we tend toward the latter view, though the algorithm reported here combines both approaches. Local segmentation operators are used to hypothesise discontinuities which are only retained if they delimit one or more describable segments . A contour is considered describable if the mean-squares residual associated with the most likely approximating primitive falls below some threshold. No attempt is made to accurately locate discontinuities, nor is a complete description required. In our view the primary goal of a bottom-up segmentation process should be to locate only those data sets which may be reliably described. This conservative approach provides an alternative to algorithms that obtain a fuller description at the cost of imposing interpretations which may not be appropriate .