Divide-and-conquer strategies for estimating multiple transparent motions

Transparent motions are additive or multiplicative superpositions of moving patterns and occur due to reflections, semi-transparencies, and partial occlusions. The estimation of transparent motions remained a challenging nonlinear problem. We here first linearize the problem in a way which makes it accessible to the known methods used for the estimation of single motions, such as structure tensor, regularization, block matching, Fourier methods, etc. We present the results for two motion layers but there is no limit to the number of layers. Finally, we present a way to categorize different transparent-motion patterns based on the rank of a generalized structure tensor.