Video object segmentation based on graph cut with dynamic shape prior constraint

In this work, we present a novel segmentation method for deformable objects in monocular videos. Firstly we introduce the dynamic shape to represent the prior knowledge about object shape deformation in a manner of auto-regressive model which treats the shape as a function of subspace shapes at previous time steps. Then both spatial-temporal image information and model prediction are fused in the framework of Markov random field energy, which can be effectively minimized by graph cut algorithm so as to achieve a global optimum segmentation. To capture model variations, both the orthogonal basis and the autoregressive model parameters are updated on-line using final segmentation results, thereby forming an effective closed loop system. Finally, promising experimental results demonstrate the potentials of the proposed segmentation method with respect to noise, clutter, and partial occlusions.