Dynamic adaptation and subdivision in 3D-SOM application to surface reconstruction

Surface reconstruction and structure representation from unorganized sample points are key problem in many applications for whose neural networks are starting a slight breakthrough. In this framework, we propose an original neural network architecture inspired from Kohonen's self-organizing maps, based on an adaptive learning process applied to a generalized mesh structure that leads to a coherent topological definition of the surface, represented by a points cloud, given as input. This representation tool seems to be efficient in most cases but some weak adaptation drawbacks appear on certain examples. We propose local neighborhood propagation and subdivision process that solves those miss-adaptation problems