Question about your paper Recursive Social Behavior Graph for Trajectory Prediction

[Xiejc97]

Hi,

Recently I read your paper "Recursive Social Behavior Graph for Trajectory Prediction", and I think it is very useful.

I want to ask when will you share the code about this paper. Besides, in your paper, you propose the Exponential L2 Loss and I have some problem about the parameter gamma. Assuming that the history length is 8 and the forecast length is 12, according to the loss function you proposed, which of the following two is the exponential term:
(1) gamma = pred_len = 12, then the exponential terms are [exp(1/12) exp(2/12) ... emp(12/12)]
(2) gamma = obs_len + pred_len = 20, then the exponential terms are [exp(9/20) exp(10/20) ... emp(20/20)]

In this project：https://github.com/ApeironY/PCCSNet/blob/main/utils/metrics.py#L31, you implemented this exponential loss function in code. In your code, gamma = a.shape[1], so the gamma is 12 and I think you implement the exp_l2_loss in the above (1).
But in your paper, you did some ablation study about the parameter gamma: 5 20 50 and then choose the best 20.

If you follow the idea of this code, how much should gamma be set when setting the parameter to other values?

[GothicAi]

Thank you for your appreciation! We will have it as a TODO to share the code. For gamma in Exponential L2 Loss, it is irrelevant with pred_len or obs_len. The terms are set as [exp(1/gamma) exp(2/gamma) ... exp(12/gamma)], where gamma can be a proper value such as 12 or 20. Hope this would be helpful.

[Xiejc97]

@GothicAi
Thanks for your answer and I have a clear understanding of the definition of index term.

But I also have some problems about the ablation study about the parameter gamma: In your paper, you say gamma = 20 is the best, but in this code gamma = 12. Does this mean that the gamma value in the paper needs to subtract obs_ length(8), and the result is the gamma value in the code?

When I do the ablation experiment on the parameter gamma, if I make gamma = 50 as in your paper, so gamma in the code is still 50, or 50-8 (obs_length) = 42?

In addition, will you consider open source for the programs in other parts of this article? I am very interested in the recursive social behavior graph in the article and want to learn it through code.

Anyway, thank you again!

Best,
Jincan

[GothicAi]

Thanks for raising this question! The best value of gamma may differ from different networks, and we use 12 in this repo while 20 for RSBG. It is not necessary to subtract obs_length.