Automatic segmentation for semantic objects from multiview images

In this paper, we present a fully automatic method to segment the semantic objects from multiview images. This method consists of two parts: firstly, the efficient segmentation initialization can be realized by automatically locating the semantic objects after analyzing the saliency map based on the visual attention model. The computation of the saliency map involves multiple sources of image information, such as depth, motion and wavelet features. Then, the binary label graph-cut based segmentation is performed to separate the semantic objects from the background. In the definition of energy function, we combine the colour with intensity clues to enhance the efficiency of the boundary term, which is efficient for the final segmentation output. The experimental results demonstrate that we can achieve a satisfying segmentation performance with low computation as well as high speed.