CamP egocentric vs. allocentric parameterization

[lxndrrss]

Hi, first of all thanks for this grat paper and code.
I have a few questions regarding the paper:

1. In the paper you mention the two possible ways of modelling the cameras extrinsic pose (egocentric vs. allocentric), however I can't seem to find which one you actually used in the end. Looking through camera_delta.py it seems to me that the residuals are modelled as egocentric parameters. Is that correct or am I misunderstanding something here?
2. As far as I understand the preconditioning matrix is derived from how much a projected points "pixel-coordinates" change based on the change of the residuals. If the points to compute this matrix are sampled from each cameras frustum, meaning that they are essentially the same for each camera in its own reference frame, and the residuals are also defined in the cameras reference frame, would that not mean that the preconditioning matrices should be more or less equal for all cameras? Assuming they have identical focal length and intrinsics of course.

I hope I am not completely misunderstanding something here. I would really appreciate if you could find the time to answer these question!