Improved prediction structure for multi-view video coding