It looks like the FaceCropAlign class does not take advantage of existing annotations, and hence, face detection is ran no matter what ...