OpenDriveLab
Why predict past trajectory in the tracker former
Thank you for your amazing work! And now I try to reproduce it in another dataset. But in data convertion stage, I noticed that there is an extra head to predict past trajectory of objects in scenario. Futhermore, there is one more past trajectory loss introduced in training, which is not shown in the paper (It also might be due to my careless reading habits, and please figure it out for me if possible, thx in advance!).
From MUTR3D, there is also no extra loss for past trajectory prediction, both in paper and code, so I`m very curious about this design, and I hope you could share your original ideas with me. Thanks anyway!
In data convertion stage, the "fut_traj" is introduced for tracking former? And I think the past trajectory is also design for tracking.
In data convertion stage, the "fut_traj" is introduced for tracking former? And I think the past trajectory is also design for tracking.
I see. So you mean the past trajectory is treated as prediction input instead of output?
I am not sure your mean about the "predict past trajectory of objects in scenario" in data convert stage due to I just found the "fut_traj". And I think the "fut_traj" is extract to supervise tracking task. Maybe this issue: https://github.com/OpenDriveLab/UniAD/issues/21 will have some cues that you want to know.