Show description

This has a double effect: it reduces the number of 3D points to be processed and it allows to filter out the noise in the data. A procedure particularly suitable to this task is Neural-Gas [69] which is an evolved version of Nearest Neighbour quantizer. It is an iterative algorithm, based on a soft-max adaptation rule, which allows to get close to the optimal solution with some reliability. This technique has been further potentiated by the Hyperbox processing procedure [70]. In this procedure, the volume occupied by the 3D data points is partitioned into a set of parallelepipeds (the hyperboxes).

IEEE Trans. on Imag. Proc. 1995; 4(10):1407-1416 49. Terzopoulos D and Waters K. Analysis and synthesis of facial image sequences using physical and anatomical models. IEEE Trans. PAM I 1993; 15(6):569-579 50. Terzopoulos D and Metaxas D. Dynamic 3D Models with Local and Global Deformations: Deformable Superquadrics. IEEE Trans. PAMI 1991; 13(7):703-714 51. Moody 1 and Darken CJ. Fast Learning in Networks of Locally-Tuned Processing Units. Neural Computation 1989; 1:281-294 52. Poggio T and Girosi F.

The drawback of HMDs is that resolution degrades rapidly by increasing the fov. Most of the HMDs have a fov in the range of 50-80 degrees which produce a tunnel like effect, that is the image is seen as it were formed at the end of a tunnel. The highest quality HMO is that of n-Vision which allows a 1280 x 960 resolution with a 50 degrees of fov monocular. This field of view can be increased when the two images are projected on the two CRT displays offset to reproduces stereo disparity. Its price is around 60,000$.