A Neural Network-Based Feedforward Architecture for Recorvering 3-D Motion Information of Curved Surfaces

Abstract

A neural network-based system for recovering 3-D motion information of curved surfaces from 2-D optical flow parameters is proposed in this paper. A feedforward network architecture that has explicit and concise physical meaning is adopted. A self tuning scheme with an unsupervised learning rule, that controls the dynamics of the system, is also employed. Moreover, a mechanism for preattentative focus, which effectively suppresses the spurious solution of the estimation, is embraced as well.