can you share the detector (PointRcnn) which get the detection results about ab3dmot

[xinshuoweng]

Hey! It has been a long time so I kind of forget the detail now. But I do remember that I have augmented the training with the tracking training dataset. The pre-trained model on OpenPCDet is for the KITTI detection benchmark so it may not be optimal for the tracking benchmark

[vkorotkine]

Hi Xinshuo,
I am a little bit confused still. I think the val dataset is the same as the training dataset for KITTI tracking, right?
Then the object detector is trained on the val set. And the AB3DMOT code, using those object detections, is evaluated on the val set.
Wouldn't this be a problem for the results? Maybe I misunderstand what is meant by KITTI val set?
I think the question of what the object detector was trained on is therefore very important for the results.
Thank you
Vassili