Joint space-time motion-based segmentation of image sequences with level set PDEs

We present a novel approach to motion segmentation by formulating it as the computation of the volume spanned by moving regions in the 3-dimensional spatio-temporal domain, in contrast to standard approaches to motion segmentation which are based on 2-dimensional spatial segmentations. We propose a Bayesian formulation for motion-based segmentation of image sequences, and we solve this Bayesian estimation problem through level set partial differential equations, which we then generalize to the case of multiple motion volumes. In addition, we provide an algorithm for estimating motion parameters independently of any segmentation. The experimental results validate our proposed approach.