Three-dimensional alpha shapes

Reviewer: Joseph J. O'Rourke

The a -shape S a of a set S of n points in Euclidean 3-space is a polytope whose boundary is a particular collection of triangles, edges, and vertices determined by the points of S . Let T be a subset of S of one, two, or three points. Then the convex hull of T , conv( T ), is part of the boundary 6 S a of the a -shape if and only if the surface of some sphere of radius a includes exactly the points of T while its interior is empty of points of S . Thus a triangle conv( T ) is part of 6 S a if and only if an open a -ball exists that can “erase” all of the triangle but leave its vertices. S a is defined for all 0? a ?? , with S 0 =S and S ? = conv S . Every edge and triangle of S a is present in the Delaunay triangulation D of S , and every edge and triangle in D is present in some S a . If a is varied continuously over its full range starting from ? , the convex hull of S is gradually eaten away by smaller and smaller a -ball erasers, eventually exposing the original set of points. In between, the <__?__Pub Fmt nolinebreak> a -shape<__?__Pub Fmt /nolinebreak> polytope bounds a subcomplex of D that represents a “concrete expression of [the] shape” of S . It has been used for cluster analysis, molecular modeling, and the analysis of medical data, among other applications. This work is one of the few papers coauthored by <__?__Pub Fmt nolinebreak>Edelsbrunner<__?__Pub Fmt /nolinebreak> that proves no theorems. Its aim is to explain the a -shape and related geometric notions for a community of potential users. The mathematical notions are presented with precision and clarity. The paper represents an unusual and welcome balance between theoretical and practical issues. Since all a -shapes are contained in the Delaunay triangulation, D can serve to store all of S a at once. The authors compute D via a simple incremental flipping algorithm. With each edge and triangle of D is associated a single <__?__Pub Fmt nolinebreak> a -interval<__?__Pub Fmt /nolinebreak><__?__Pub Caret1> representing the values of a for which that simplex is present in S a . The elementary but nontrivial geometric computations necessary for implementing the required calculations are explained carefully. They are all expressed in terms of determinant evaluations. Degenerate cases, which often plague geometric algorithms, are handled by the “simulation of simplicity” method pioneered by the authors. This method breaks degeneracies symbolically in a uniform manner, resulting in robust code. The authors have implemented their algorithm, packaged it in an attractive SGI interface, and distributed it freely and widely via ftp. Analyses of runtimes are presented and contrasted with the worst-case O n 2 bound; not surprisingly, most data sets do not exhibit worst-case growth. This superb paper can stand as a model of how to bridge the gap between theory and practice in computational <__?__Pub Fmt nolinebreak>geometry.<__?__Pub Fmt /nolinebreak>