Improving segmentation results by studying surface continuity

Abstract

This paper presents a process to improve the quality of range image segmentation by using geometrical relationships. The proposed technique consists of studying the surface continuity of an automatically generated surface model. Generally, surfaces are extracted independently (e.g., by means of a region growing algorithm) thus information about their connectivity is lost. Assuming that in most of the cases a surface cannot be disconnected with the others present in the given scene, occluded areas and crease edges can be recovered. Occluded regions are recovered by connecting surfaces that are represented by the same parameters. In addition, enforcing geometrical constraints, such as surface intersections, crease edges are recovered improving significantly the final model. Experimental results with automatically segmented real range images are presented.