An EM-like Algorithm for Motion Segmentation via Eigendecomposition

This paper presents an iterative maximum likelihood framework for motion segmentation via the pairwise checking of pixel blocks. We commence from a characterisation of the motion blocks in terms of a matrix of pairwise similarity weghts for their motion vectors. The eigenvectors of this similarity weight matrix represent the initial pairwise clusters, i.e the independant motions present in the scene. We develop a maximum likelihood framework which allows to update both the link weight matrix and the associated set of pairwise clusters. We experiment with the resulting clustering method on a number of real world motion sequences. Here ground truth data indicates that the method can result in motion classification errors as low as 3%.