Implicit Representations for Recovery and Reconstruction of Motion Layers

In this paper, we propose a level set approach to decompose the image plane into segments with consistent apparent motion and recover parametric models that best describe such a motion. The number of motion planes and initial motion parameters are obtained by robust block-wise optical flow estimation, followed by non-parametric clustering in the motion parameter space. Implicit representations are used to further improve the estimation and to segment the image plane according to consistent motion. Our approach integrates a smoothness constraint, a spatial-temporal motion constraint and a color-driven segmentation module within the multiphase level set framework. Therefore, our method is able to recover the different topology of the motion layers and provide smooth motion segmentation that is consistent with the geometry of the scene. Experimental results on both natural and synthetic image sequences demonstrate the potential of the proposed technique.