Motion Estimation via Scale-Space in Unsupervised Deep Learning

We present a potential application of the conventional scale-space theory to the estimation of optical flow in the deep learning framework. An unsupervised learning scheme for the computation of optical flow is integrated with a Gaussian scale space. The hierarchical propagation of intermediate estimations via a consecutive scales demonstrates a potential in the course of optimization leading to a better local minimum. The landscape of loss function associated with an optical flow problem in a neural network framework is highly complex and non-convex, which requires to guild the optimization path in such a way that a solution at a plateau region. The qualitative comparison of the optical flow solutions via a Gaussian scale-space provides the characteristics of solutions at different scales, thus provides a way to take into consideration of scales in further improving accuracy.