I have tested many scenes, and the pose of the camera is a little different between Registration Output and Metadata(XMP) Output. The Registration Output, of which the re-projection error is smaller, is frequently better than Metadata.
I used the formula t = - RC , and found there is some difference between t_reg(exported by Registration) and t_meta(exported by Registration) .
What is the real meaning of Rotation and Position in Metadata(XMP)?