Can Scale-Consistent Monocular Depth Be Learned in a Self-Supervised Scale-Invariant Manner?