3D object modeling with graphics hardware acceleration and unsupervised neural networks

Abstract

This paper presents a methodology for reaching higher performances when modeling 3D virtualized reality objects using Self-Organizing Maps (SOM) and Neural Gas Networks (NGN). Our aim is to improve the training speed of unsupervised neural networks when modeling 3D objects using a parallel implementation in a Graphic Process Unit (GPU). Experimental tests were performed over several virtualized reality objects as phantom brain tumors, archaeological items, faces and fruits. In this research, the classic SOM and NGN algorithms were adapted to the data-parallel GPU, and were compared to a similar implementation in an only-CPU platform. We present evidence that rates NGN as a better neural architecture, in quality terms, compared to SOM in the task of 3D object modeling. In order to combine the NGN accuracy with the SOM faster training, we propose and implement a hybrid neural network based on NGN using SOM as seed. Our experimental results show a considerable reduction in the training time without affecting the representation accuracy. © 2011 Springer-Verlag.