As to my current understanding, Vision API does support yaw and roll values from a detected face. However, it only supports a discrete set of values such as pi/4, pi/2, and etc.