Near Real-Time Motion Segmentation Using Graph Cuts

We present a new approach to integrated motion estimation and segmentation by combining methods from discrete and continuous optimization. The velocity of each of a set of regions is modeled as a Gaussian-distributed random variable and motion models and segmentation are obtained by alternated maximization of a Bayesian a-posteriori probability. We show that for fixed segmentation the model parameters are given by a closed-form solution. Given the velocities, the segmentation is in turn determined using graph cuts which allows a globally optimal solution in the case of two regions. Consequently, there is no contour evolution based on differential increments as for example in level set methods. Experimental results on synthetic and real data show that good segmentations are obtained at speeds close to real-time.

• Type: Conference
• Year: 2006
• Where: DAGM
• Authors: Thomas Schoenemann, Daniel Cremers
• Keywords: Bayesian A-posteriori Probability, DAGM 2006, Image Processing, Integrated Motion Estimation, Level Set Methods