Regarding the confusion over inconsistent SocialCircle results presented in different articles

[18821252916]

I'm very interested in your work! But，I have noticed that the results presented by different papers regarding the performance of SocialCircle and E-V^2Net on the univ dataset are inconsistent. In the RE paper, it is shown that the results of the SocialCircle and E-V^2Net model are 0.23/0.42, while in the original paper of SocialCircle, the results are both 0.20/0.34. Why is this so? However, the performance on other datasets is consistent between the two papers.

[cocoon2wong]

Hi @18821252916

Thanks for your questions!

This difference is due to the fact that the original use of univ differs from other baselines:

• The univ used in the two papers you mentioned with better results included only univ1 as the test set;
• The corrected univ includes univ1 and univ3 as test sets (tho it is still called the leave-one-out).

These results were modified in the journal versions of these two papers, “Another vertical view: A hierarchical network for
heterogeneous trajectory prediction via spectrums” and “Socialcircle+: Learning the angle-based conditioned interaction
representation for pedestrian trajectory prediction” respectively, and we have also added a description of this modification to the README in their repos like SocialCircle.

As a distinguishing mark, we use univ13 as a suffix for these corrected weights, and you can see this difference in the weights we have provided.

Thank you!