Wyner-Ziv coding of multiview images with unsupervised learning of two disparities