Surface Reconstruction Using Neural Networks

Surface Reconstruction from Unorganized Points Using Kohonen's Self-Organizing Map

Yizhou Yu
Computer Science Division
University of California at Berkeley

We introduce a novel technique for surface reconstruction from unorganized points by applying Kohonen's self-organizing map. The topology of the surface is predetermined, and a neural network learning algorithm is carried out to obtain correct 3D coordinates at each vertex of the surface. Edge swap and multiresolution learning are proposed to make the algorithm more effective and more efficient. The whole algorithm is very simple to implement. Experimental results have shown our techniques are successful.

Paper postscript and pdf files as it appears in IEEE Visualization'99 Late Breaking Hot Topics.

The original point set of the Stanford Bunny and three intermediate reconstructions at three levels of resolution.

Two views of the final reconstruction for the Standford Bunny.

The original point set of an open surface generated from a mixture of Gaussians and the texture-mapped reconstructed surface.

The original point set and reconstructed surface of a mannequin.

(Back to Yizhou's Home Page)