Object part descriptions are important because they reflect natural structures in the real world and support efficient object recognition. Part segmentation is the crucial step for building such a description. In this research, we propose a novel approach to part segmentation using multiview range images. Object data obtained from multiple viewpoints contain much more shape information about a complicated object than data taken from a single view. Motivated by physics, we compute the simulated charge density distribution over an object surface which is tessellated by a triangular mesh. We then detect deep surface concavities by locating local charge density minima. The object is then segmented into parts at these surface concavities. Although this approach deals with a complete 3D object, it performs the reconstruction only on the object surface, rather than its interior as would the commonly used voxel-based approaches. Thus, it requires much less unknowns. Our method can also be extended to perform part segmentation using single-view range data and can be generalized to take any polygonal mesh as input. A significant aspect of this work is that it computes surface features using global, rather than local, data. In addition, it avoids explicitly dealing with the scale problem, a very difficult issue in the usual curvature-based approaches. The feature that distinguishes this approach from others is the application of simulated charge density distributions. The results can also be used for performing other tasks, such as edge and corner detection.
K. Wu, M.D. Levine
Object segmentation and representation. (1) Object data integrated from range images in four views. (b) The triangulation of the object surface. (c) The simulated charge density distribution. Charge densities shown as gray levels normalised between 0 and 255. (d) Two segmented parts. (e) The object model represented by parametric geons.