Closed jun-long-li closed 2 years ago
Hi Jlong
Sorry for the delay. As we have talked in emails, now I have free time to get back to you.
I just update the model weights and expert-goal examples for these two subsets. Can you give it a try on Zara2 & Univ? I reproduce the results for these two subsets using the k-mean retrieval fashion.
Note that with the k-mean fashion, I find the results for Zara2 now is ~0.9/0.26, which is different from 0.12/0.24 in the paper. I use 20 centroids for k-mean on top of DTW (100 smallest results) this time, whereas the original paper just use plain DTW. Looks like it hurts the FDE but benefits the ADE a bit.
For the Univ subset, I find the updated result ok.
Thank you again for your interest. Please let me know if any problem.
Thanks a lot! With these, I can reproduce similar results now.
Hi, I can't find the corresponding files for zara2&univ in '_/checkpointethucy'. Would you please provide it for better reproductivity?
Thanks