Performance Gap on Argoverse 2 Validation Set

[Yipeng-Zhou]

Thanks for sharing this perfect work!

After evaluating the pre-trained model on Argoverse 2 validation set using the default configuration, I found a performance gap between the results I got and the data in the table IV (the last row). Here is the comparison:

This issue does not occur on the Argoverse 1 validation set.:

Looking forward to your reply!

[MasterIzumi]

@Yipeng-Zhou Hi, thanks for your feedback! Actually, the released checkpoint for AV2 is not the one we used in the paper. I will check this issue and upload the correct one soon.

[MasterIzumi]

Sorry for the late reply. Unfortunately, I cannot find the exact checkpoint for the experiment. So I re-train the network and get a checkpoint that achieves similar performance. I have updated the weight at saved_models/simpl_av2_bezier_ckpt-fix.tar, and the quantitative result on AV2 validation set is:

Validation set finish, cost 72.22 secs
-- minade_1: 2.045 minfde_1: 5.594 mr_1: 0.663 brier_fde_1: 5.594 minade_k: 0.777 minfde_k: 1.452 mr_k: 0.196 brier_fde_k: 2.069

If you have any questions, please let me know.
Again, thanks for your valuable feedback!