MPI-I-2003-4-007

We propose a new surface reconstruction algorithm based on an incrementally expanding neural network known as Growing Cell Structure. The neural network learns a probability space, which represents the surface for reconstruction, through a competitive learning process. The topology is learned through statistics based operations which create boundaries and merge them to create handles. We study the algorithm theoretically, calculating its complexity, using probabilistic arguments to find relationships between the parameters, and finally, running statistical experiments to optimize the parameters.