Ground truth projection problem

[jiheddachraoui]

Hello, thank you for your interesting work.
I am working on camera–LiDAR registration for pose estimation. While using the provided synchronization code and tutorials, I observed that in some cases (particularly when the vehicle is turning), the ground truth projection onto the camera image looks worse than the predicted projection.

Additionally, I would like to verify the quality of the provided ground-truth poses, since this issue may also contribute to inconsistent numerical results (pose errors) even though the visual alignment should indicate no error.

[keenan-burnett]

Hello,

For the boreas-objects-v1 subset, we had a slightly different setup than the rest of the dataset. There may be some temporal synchronization issues. It's possible that the ground truth poses for this subset may have some issues. It's interesting because the labels seem to project okay into the radar and lidar data, so that leads me to suspect temporal synch or extrinsic / intrinsic calibration issues for the camera images. Unfortunately, I no longer have the bandwidth to address this issue. If you find a reasonable fix, please let me know and I can merge your proposed fix into the dataset / repo.

The rest of the dataset (which we use for mapping and localization research) has been more thoroughly vetted for temporal synch, intrinsic cal, extrinsic cal, etc. Although, I understand that may not be of much use to you since that portion of the dataset is not labelled.

[keenan-burnett]

Also, this seems like a duplicate of #64

[jiheddachraoui]

I am working with the localization dataset (not the object detection one) and have encountered similar issues in pose estimation. Specifically, I am using camera–LiDAR projection to estimate the relative pose and compare it with the provided ground truth. I may not be following the best practices, so if you are confident about the data quality, I would greatly appreciate any guidance on the recommended practices. I have already carried out several tests, as mentioned in my initial comment

[keenan-burnett]

Hello, if you can post some screenshots of what issues you've encountered, that would be helpful.

Also, some code snippets or pseudo-code of your approach.

There's not enough information here for me to be able to comment on what the potential issue may be.