History: Márques's Mestrado Thesis


Introduction

Methods for reconstructing from range images a surface of a three dimensional object have evolved rapidly in the last years. One reconstruction paradigm, known as a local deformation one, is to reshape dynamic and locally a geometrically simple, but closed mesh until it fits the range data with specified accuracy. To our knowledge, the proposed deformables algorithms until today, can only produce meshes which are topologically equivalent to the initial one.

Radial Flow Model (RFM) is a method of reconstruction of a surface from a unregistered range image  that uses this deformable principle. It was created under the inspiration of the Chen and Medioni's work [1], by combining  advantages of its and of a  function graph-model [2],. However, it overcomes their limitations such as self-intersections, restriction to surfaces with number of genus zero in the first approach and cracks in the reconstructed surfaces in the second approach. RFM generates a 3D closed mesh without holes, without self-intersections  and without restriction to topological class of the surface to be reconstructed.

The basic principle of RFM is simple,  it starts by making a radial mapping of the vertices of a decahedron to sampled points, and as they grow towards their corresponding points,  growing mesh is refined by genering new vertices and new radial mappings until that the mesh gradually adjusts in the range image. In order to maintain the topological consistency between the surface sampled and surface reconstructed, the topological genus estimated from range image is reflected on the mesh during the growing.
 
 

The method

From the range data the initial reference system is determined and an initial closed mesh (decahedron) is placed on the origin of this reference system. Under the influence of a radial force, the mesh inflates as its vertices move towards the corresponding points in the range image. When the length of an edge in the mesh is greater than a specified tolerance, the edge is subdivided and the mesh must be refined maintain a triangular topology.
The sequence of operations refining-inflating is repeated, until no more subdivision and no more inflation is possible. Then, reconstruction errors are evaluated. If there are no faces in the mesh i with reconstruction errors, we consider that the aimed reconstructed 3D model is achieved. Otherwise, the
triangular faces that still have reconstruction errors are grouped into disjoint sets of faces. These sets of faces are denoted fronts (of growth). From each of them we determine a new radial reference system before carrying out again the cyclic sequence refine-inflating to yield a new mesh i+1.

Moreover, whatever the surface to be reconstructed presents genus number different from zero, our method is able to realize topological surgeries during the deformation process in order to match the topology of the deformable model with the topology of the reconstructed surface.

Some reconstructed surfaces with their respective range images are showed now

Implementation

Radial Flow Model has been implemented in language C in the platform UNIX. It is executable in SUN-SPARC and PC (Linux).
For visualization of reconstructed surfaces have been used the grafic library MESA (clone freeware of grafic interface OpenGL) and the grafic package Geomview.

The implementation so depends on a topological structure management library, TDM [1], written by Prof. Ting.
 

References

[1] Y.Chen and G.Medioni,  Description of complex objects from multiple range images using an inflating ballon model , Computer Vision and  Image Understanding, Vol.61, No.3, pages 325-334,1995.

[2] H.T.Tanaka, Accuracy-based sampling and reconstruction with adaptative meshes for parallel hierarchical triangulation, Computer Vision and Image Understanding, Vol.61, No.3, pages 335-350,1995.

[3] Msu/wsu range database: Usc image (composer), http://www.cs.cmu.edu/afs/cs/project/cil/ftp/html/vision.html

[4] Wu, Shin-Ting, Topologie von Hybriden Objekten, Technische Hochschule Darmstadt, Darmstadt, Germany, February 1991, Phd Thesis.