Towards Learning Structure Via Consensus For Face Segmentation And Parsing