A simple feedforward neural network architecture for 3-D motion and structure estimation
Yi Sun, Mohamed M. Bayoumi · 2002
A simple feedforward neural network system which is especially designed to tackle the problem of 3-D motion and structure parameter estimation from 2-D optical flow parameters has been proposed. Each node and the weight of every connection adopted in the network has its explicit physical meaning. The network embraces a self tuning scheme with an unsupervised learning rule to control the dynamics of the system. It also adopts a mechanism for preattentative focus that effectively suppresses the spurious solution of the estimation.