RobotCar Seasons: inaccurate rig extrinsics

[sarlinpe]

The transformations between the rear, right, and left cameras are provided in extrinsics/*_extrinsics.txt. They are not consistent with the relative transformations that can be estimated from the absolute poses of the reference sequence overcast-summer. This negatively impacts the accuracy of multi-camera localization on RobotCar.

Additionally, the relative transformations are not even constant in the reference model - the rig constraint was not enforced when constructing the model. Nevertheless, reestimating the relative transformations from the reference model (as least-squares) gives better results, which shows that the provided extrinsics are very inaccurate.

[tsattler]

If I remember correctly, we enforced the rig constraint when computing the reference model in Colmap. The rig bundle adjuster was experimental by that time, so some things might not have worked as expected, e.g., the fact that the model is constructed from individual locations might have created some problems there.

The provided rig extrinsics are, to the best of my knowledge, identical to those provided by the original RobotCar dataset. They should have been used to generate the reference poses for the queries.

I believe that RobotCar, similar to Aachen, needs a revision of the reference poses as I believe we are running into the accuracy of the reference poses. This is interesting behavior, because by then we believed that the RobotCar dataset was actually one of the more accurate ones.
We will leave the current version of the RobotCar dataset as is for now. If you are interested in helping with a newer version of the RobotCar dataset (I currently only have limited capacities), then please ping me per email.