Confusion regarding Figure 9 in the paper

[haoningc]

Hello, I would like to determine whether the precise pose of the reference image and the true pose of the query image are obtained through the sfm method? Or is it the pose carried by the dataset itself? If the pose is obtained by the sfm method, is it unfair in Figure 9 of the paper? Because the 3D points obtained by sfm themselves also have certain errors, when the poses and 3D points obtained by sfm are back-projected onto the image, they should naturally be more accurate than the pose back-projection using the RobotCar dataset. If I make any mistakes in my expression, please forgive my ignorance.