search

ArashJavan / DeepLIO

Deep Lidar Inertial Odometry
Apache License 2.0
53 stars 10 forks source link

Regarding the ground truth #18

Open KleinYuan opened 3 years ago

KleinYuan commented 3 years ago

Hi,

I noticed that you use the sync-ed SE(3) as the ground truth of odometry. But from what I read on KITTI website, the only difference between sync and unsync oxts is just the temporal synchronization. For the odometry ground truth, shouldn't we use the odometry datasets: http://www.cvlibs.net/datasets/kitti/eval_odometry.php instead?

Otherwise, it seems the model is learning synchronized but biased SE(3)

ArashJavan commented 3 years ago

@KleinYuan unfortunately the lack of synchronization is not the only thing, also the odometry benchmark (or synced) does not contain IMU values. But the timestamps are defined from the same environment so they are relative to each other well defined. So we can relate them as done here.

Regards Arash

KleinYuan commented 3 years ago

Thanks for the reply.

  1. Could you elaborate on "unfortunately the lack of synchronization is not the only thing" ? From what I read on their websites/devkits and paper, it seems the temporal sync is the only diff between synced and unsynced oxts.
  2. In your paper, you said you are using "KITTI Odometry datasets" while in your code, you are constructing the ground truth with the synced raw datasets. Are you suggesting that they are identical?