Able to get good/comparable results for SDD in Ours-S-TT configuration about human-path-prediction HOT 5 CLOSED

Hi,

That is correct -- it's averaged during test so the signal on the test/val set during training may be slightly noisy. Correct me if I'm wrong, but it seems as though you are using the TT (sampling with >1 SD at test time which is the default we had set for K=20); or did you change the sampling SD to 1?

from human-path-prediction.

HI @HarshayuGirase

Apologies for the late reply.

I did not change the number of samples to 1. It is still K=20. The scale used is 1.3 (as in your original implementation)

Regards
Kapil

from human-path-prediction.

Hi @ksachdeva,

Thanks for clarifying. It seems as though you are using the TT, as the SD is 1.3 (different than what was used in training). Please see the Truncation Trick in section 3.1. For larger sample sizes (K=20), instead of truncating, we use a larger SD for better sample diversity, so this model you experimented with would just be O-S. Setting the SD to 1 during test time would correspond to the O-S-TT ablation.

from human-path-prediction.

Thanks @HarshayuGirase

It makes sense now. I would try again with SD = 1 during test.

as the SD is 1.3 (different than what was used in training).

I got 1.3 from your configuration file. I am assuming it is valid for SDD, was it different for ETH ?

Regards
Kapil

from human-path-prediction.

No problem @ksachdeva !

Yes, our default parameters in config file were the best parameters we found for K=20 (so it includes social + changing SD trick). I believe it was the same for ETH/UCY. I'll close this issue for now -- feel free to open another issue for other questions!

from human-path-prediction.