Result on the Argoverse 2 leaderboard

[yanconglin]

Hi, thank you for releasing the codebase. I was wondering if you could kindly clarify the following questions?

1. it seems the leaderboard result is much better than the reported values in the paper (e.g., EPE-3way). would you mind sharing some info on how you made this happen?
2. I have found several checkpoints. One is available at deflow discussion on leaderboard result , denoted as "seflow_official", and another one is refered to as "seflow_best" available at seflow_repo. Could you pleae let me know the differences between these two?
3. If I were about to make a comparison, let us say using the 1x setup without extra data, which one is to use?

I look forward to your reply. Many thanks in advance.

[Kin-Zhang]

Hi, there is a note in the README: https://github.com/KTH-RPL/SeFlow?tab=readme-ov-file#other-benchmark-models

• The paper result is under 2e-6 and epoch 50 to have fair ablation study with same setup on ZeroFlow since we directly use their weight.
• seflow_official is the weight I used in the paper result and leaderboard v1 result (since when I'm writing the paper, leaderboard v2 hasn't been published yet).
• seflow_best is the weight I trained with the note said and the result in leaderboard version 2 (we also mentioned this in updated supplementary material.) config: 2e-4 and epoch 20. You can find that also using torch.load(model), all configs are saved in the model weight. the v2 leaderboard discussion result is here: https://github.com/KTH-RPL/DeFlow/discussions/6
• Personally, I will say use the result on the leaderboard and the file name is seflow_best.ckpt.

[yanconglin]

Thank you for the clarification. Fair points. One additional question: does the best result on benchmark follow a 1x or 2x setup?

[Kin-Zhang]

1x dataset size for all results in both leaderboards (v1, v2).