Output details

Segmentation of discrete point clouds using an extensible set of templates

One of the effective techniques in computer art and computer-aided design applications is reverse engineering, which means reconstruction of geometric models from point clouds obtained by scanning physical objects such as clay figurines or mechanical parts. A typical approach is to reconstruct a polygonal mesh or a parametric surface from a point cloud. However, these models are hard to modify and to assign high-level parameters for further modifications. Segmentation is the first step of reverse engineering, where high-level primitive shapes are detected in the cloud such as planes, spheres, cylinders and others. The existing segmentation techniques do it locally such that a large sphere present in the object can be detected a number of different spheres present in different parts of the object. Another problem with the existing approaches is that they are limited in the types of primitives they can recover from scans.

Originality

We have presented in this paper a new algorithm for segmenting a point cloud. We first describe template primitives appropriate for the particular application area. Then the algorithm works by iteratively trying to fit template primitives to the input point cloud and by removing at each iteration points matching the optimal fitted primitive from the cloud.

Significance

We showed that this algorithm is extensible by various types of primitives depending on the application area. An example is given of a lacquer ware pot object processed using a non-standard super-ellipsoid primitive as one of its template primitives. Therefore, the significance of the contribution lies within the domain specific segmentation rather than uniform segmentation independent of the area.

Rigour

We thoroughly tested the algorithm on artificial objects, standard objects and scanned data. We have also demonstrated the sensitivity of the algorithm to different parameters as well as its behaviour when applied to noisy input data.