Neural-network-based models of 3-D objects for virtualized reality: a comparative study

The paper presents a comprehensive analysis and comparison of the representational capabilities of three neural architectures for three-dimensional (3-D) object representation in terms of purpose, computational cost, complexity, conformance and convenience, ease of manipulation, and potential applications in the context of virtualized reality. Starting from a pointcloud that embeds the shape of the object to be modeled, a volumetric representation is obtained using a multilayer feedforward neural network (MLFFNN) or a surface representation using either the self-organizing map (SOM) or the neural gas network. The representation provided by the neural networks (NNs) is simple, compact, and accurate. The models can be easily transformed in size, position, and shape. Some potential applications of the presented architectures in the context of virtualized reality are for the modeling of set operations and object morphing, for the detection of object collision, and for object recognition, object motion estimation, and segmentation.