Visual Object Segmentation Based On Mixture Modeling On Key Frames

In this paper, we reviewed and analyzed the related work of video object segmentation problem, and noticed that the previous work could not handle the problem of overlapping feature between foreground object and background. To address this difficulty, in this paper ,a sophisticated mixture model coupling strong image classifiers with latent spatial configuration is proposed. Different from existing video segmentation techniques which estimate a simple parametric color distribution without considering the location information, our method use both color pixel feature and histogram of color and oriented gradient, furthermore, in the process of Expectation Maximization, we apply discriminative learning method to train a series of classifiers, which has different weight on different domain of the image, governed by a location related latent variable. The proposed framework take advantage of the strength of discriminative learning algorithm, and is able to handle the challenge of overlapping feature, blur boundary and large displacement between successive frames. In order to adapt the learned model to successive frames which are not labeled by user, a feature tracking based method is used to update the location related model parameters. Furthermore, we introduce the max-flow/min-cut based Graph Cuts algorithms to optimize the segmentation result. We test our algorithm on a wide range of data and the experimental results show that our algorithm can get better results than existing methods.