Variational Shape Approximation

David Cohen-Steiner, Pierre Alliez, Mathieu Desbrun

Proceedings of the 31st Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH)

to appear in 2004




Motivation

Many 3D datasets, especially scanned meshes, have far more points, edges and triangles than are needed to represent their geometry. A central problem in computational geometry is to be able to simplify excessively large meshes: find a concise but geometrically accurate representation of such meshes. The problem has numerous applications ranging from geometry compression to reverse engineering.

Goal of This Research

The authors wish to answer the following questions:
  1. Given a 3D surface, a target number of face elements, and an error metric, what is the best geometric approximation of the object that one can find with this face budget?
  2. Given a distortion tolerance, what is the smallest polygonal mesh approximant with a distortion less than the tolerance?
This problem is NP-hard so researchers shy away from it.

Goal of This Paper

This paper casts the shape approximation problem as a variational geometric partitioning problem (i.e. as the problem of finding a partition that minimizes a given error metric). Using the concept of geometric proxies, the algorithm drives down the distortion error by repeatedly clustering faces into best-fitting regions. This paper also presents a new metric based on normal deviation.

Related Work

A piecewise linear approximation of the geometry of a mesh with optimal efficiency appears to be theoretically difficult.

Results

Advantages of the presented shape approximation approach include:
  1. Symmetries are quickly found
  2. Anisotropy is automatically detected and exploited
  3. Regions line up with the features
  4. Although the simplified meshes do not have nicely shaped triangles with smooth sampling gradation, efficiency is gained by respecting features and symmetries
Disadvantages of the approach include:
  1. Algorithm is slower than greedy methods
  2. Algorithm only handles piecewise linear 2-manifolds, although extension to point clouds is feasible
  3. Qslim (cite garlandAndHeckbert98) outperforms algorithm if a given number of triangles is sought, since algorithm is not error-driven and is designed for polygonal outputs
  4. Mesh not allowed to be non-manifold
Possible improvements to the algorithm include allowing the output mesh to be non-manifold and locally extracting a dual mesh so that the output mesh will have nicely shaped triangles in round regions and elongated triangles in anisotropic regions. Other future work includes trying a Sobolev metric, simplifying large scientific sequences using variatonal motion approximation with a 4D (space + time) metric, geometry compression, higher-order proxies, and the notion of shape shape complexity and how it relates to the choice of a metric towards a sampling theory for shapes.

Bibliography



ctj