Optical flow estimation via neural singular value decomposition learning

Abstract

In the recent contribution [9], it was given a unified view of four neural-network-learning-based singular-value-decomposition algorithms, along with some analytical results that characterize their behavior. In the mentioned paper, no attention was paid to the specific integration of the learning equations which appear under the form of first-order matrix-type ordinary differential equations on the orthogonal group or on the Stiefel manifold. The aim of the present paper is to consider a suitable integration method, based on mathematical geometric integration theory. The obtained algorithm is applied to optical flow computation for motion estimation in image sequences. © Springer-Verlag 2004.