Learning geometric consistency and discrepancy for category-level 6D object pose estimation from point clouds.