Optical Flow Estimation via Neural Singular Value Decomposition

In the recent contribution [9], it was given a unied 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 specic integration of the learning equations which appear under the form of rst-order matrix-type ordinary dierential 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 al-gorithm is applied to optical flow computation for motion estimation in image sequences.