Towards Learning Structure via Consensus for Face Segmentation and Parsing