We introduce a noise-resistant algorithm for reconstructing a watertight surface from point cloud data.

It forms a Delaunay tetrahedralization, then uses a variant of spectral graph partitioning to decide whether each

tetrahedron is inside or outside the original object. The reconstructed surface triangulation is the set of triangular

faces where inside and outside tetrahedra meet. Because the spectral partitioner makes local decisions based on

a global view of the model, it can ignore outliers, patch holes and undersampled regions, and surmount ambiguity

due to measurement errors. Our algorithm can optionally produce a manifold surface. We present empirical

evidence that our implementation is substantially more robust than several closely related surface reconstruction

programs.