search

Kay1794 / Aware-of-the-history

Official PyTorch code for our ECCV'22 paper Collaborative Uncertainty in Multi-Agent Trajectory Forecasting.
16 stars 2 forks source link

Reproduce results & local behavior data #3

Open josephinemonica opened 1 year ago

josephinemonica commented 1 year ago

Thank you for the interesting work.

We want to use your work, and trained the lanegcn + LBA network on the Argoverse data (using the Local behavior data you provided in github), but we're not able to reproduce the results (as in the paper). The results that we got is much worse, even worse than the baseline that do not use the local behavior data. We tried training it twice, both results are consistently worse. Can you tell us how to reproduce the results?

We are wondering if the problem is in the local behavior data. We checked the local behavior data (from link you provided https://drive.google.com/file/d/1GmnCd4ByA9s6b9RKgQHD3wipQn-MP91u/view) It does not look like the data you showed in the paper. Can you please double check the data/ publish the code to extract the local behavior data?

Kay1794 commented 1 year ago

Hi @josephinemonica, Thank you very much for your interest. Sorry to hear that the code may have a reproduction issue...Sure, I will look into the data. Meanwhile, have you tried visualizing local behavior data as well as the agent's trajectory of any of the training samples? Also, would you mind sending me a log file generated during the training through email(to yiqizhon@usc.edu)?

josephinemonica commented 1 year ago

Thanks for the reply. I tried to visualize the local behavior data for a scene, but it does not look like illustration you showed in the paper. Specifically, the local behavior are far (>1m) from the agents' current location

josephinemonica commented 1 year ago

I sent the log file to your email. Thanks a lot

josephinemonica commented 1 year ago

Hi, I checked again the local behavior data is fine I think, just some scenes do not have very good data. But overall okay. But I can not reproduce your results. The results with LBA is significantly worse than reported in the paper. Can you help? Thank you.