get inaccurate camera extrinsics

[pulotaha]

Thank you for your excellent work!

I am having some problems when reproducing the Estimating Camera Extrinsics part of your paper.

I have set the intrinsics matrix to [[4.2647, 0, 0.5],[0, 4.2647, 0.5],[0, 0, 1]].
Then use cv2.solvePnP to get the rotation vector and t vector.
But the camera is not oriented towards the coordinate center (i.e., the car's center of mass) after being converted to an extrinsics matrix visualization.

May I ask how you used PnP to get the extrinsics parameters?
Is there anything wrong with my operation?

Thank you for your help.

[nviolante25]

Hi, thank you for your interest in our work!

I have set the intrinsics matrix to [[4.2647, 0, 0.5],[0, 4.2647, 0.5],[0, 0, 1]].
Just to clarify, did you do that here?

Which tool are you using for the visualization? OpenCV's PnP convention is different that the convention used for training the model. We extracted the poses using OpenCV's convention and then transformed them for training. We provided this script to visualize poses.

Hope it helps!