Why are results on KITTI MOT Validation set in README different from those in your paper? about ab3dmot

[xinshuoweng]

In the KITTI dataset, the 3D bounding box ground truth is not provided for DontCare objects, so we cannot use 3D IOU to check for donecare objects in KITTI because doing so will essentially cause unnecessary false positives. As a result, change to 2D you to check could get better but also more reasonable numbers.

The paper is not finalized yet. I am preparing for the camera-ready version, which will ultimately have the current reasonable numbers shown in this repo.

[xinshuoweng]

However, in the future, if a 3D tracking dataset provides us the 3D bonding box ground truth for DontCare objects, we should change the evaluation code to use 3D IOU to check.