Automatic Object Cosegmentation in Sparse Multiview Images

In this paper, an automatic approach for object co segmentation in sparse multiview images is presented. To extract the foreground object in multiview images with wide baselines, the cues of color and geometry are combined in our framework of double iterative loops. In the outer loop, we iteratively apply geometric background extraction and iterated graph cuts. Based on previous silhouettes, we effectively exploit the epipolar tangency constraint to extract the background in each image. After learning the bilayer color models, iterated graph cuts are applied to obtain the refined silhouettes, which are further employed in the next round of outer loop. In the inner loop, iterated graph cuts, due to the pixels in geometric extracted background are often far from the target object, a background gain factor is proposed to enhance the power of the background color model. In our method the 3D model of the object or the depth maps are discarded for limited sparse multiview images. Experiments show that based on a small number of multiview images our double-loops system can effectively extract the object.