Motion segmentation via synchronization

In this paper we consider the problem of segmenting points in a collection of images that contain multiple moving objects. Our contribution is three-fold: (i) we propose a matrix representation of segmentation that permits to formulate the problem in terms of 'synchronization' of binary matrices; (ii) we derive a spectral solution to solve such a problem, which is inspired by previous works on synchronization of rotations, homographies, rigid motions and permutations; (iii) we explain how our solution can be interpreted in terms of spectral clustering. The proposed approach is validated on both synthetic and real scenarios, in addition to the Hopkins benchmark.