Memory Efficient Surface Reconstruction Based on Self Organising Maps

Abstract

We propose a memory efficient, scalable surface reconstruction algorithm based on self organising maps (SOMs). Following previous approaches to SOM based implicit surface reconstruction, the proposed SOM has the geometry of a regular grid and is trained with point samples extracted along the normals of the input data. The layer by layer training of the SOM makes the algorithm memory efficient and scalable as at no stage there is need to hold the entire SOM in memory. Experiments show that the proposed algorithm can support the training of the very large SOMs that are needed for richly detailed surface reconstructions.