Figure 2: Enhanced reconstruction algorithm. (a)
Points extracted from CT volume
via isosurfacing. (b) 
-solid of points in (a). (c) Reduced
mesh. (d) Reconstructed model of the tibia. Curves highlight
patch
boundaries.
Figure 3: Reconstructed knee model of the
Visible Male, showing femur,
tibia, fibula and patella. (a) Polyhedral model approximating
the
extracted surface from the CT volume. (b) Smooth surface
patches
model of (a), with curves highlighting patch boundaries.
Some objects have sharp geometric features such as corners and
creases. Cubic A-patches can be joined to achieve surface 
continuity, but the continuity constraints can be relaxed along
selected edges or faces to permit the representation of sharp
corners,
rectilinear or curvilinear edges, and flat
faces [2]. These
features must however be completely
contained in, or coincident with, a collection of edges or faces
of
the supporting mesh.
We start by extracting, as in our previous algorithm, a two-manifold from the Delaunay triangulation of P. This fine mesh can be used to estimate dihedral angles between adjacent triangles and detect and tag sharp features. A mesh decimation algorithm is then applied to reduce the number of triangles, especially in areas of small curvature. The reduced mesh is used as the base for a simplicial-hull construction. The result is a mesh of tetrahedra surrounding the triangulated two-manifold, that will act as support for the surface patches.
At this points, weights for all patches are computed from
estimated
surface normals and least-squares approximation of data points,
starting with faces containing sharp features. We obtain in this
way a
mesh of surface patches that are -continuous, except along
selected features, and approximate the points. Finally, we run an
energy minimization algorithm to optimize distribution of
curvature
and improve data fitting.
Figures 2 and 3(b) show examples of reconstruction using this algorithm.
