search

yuantianyuan01 / StreamMapNet

GNU General Public License v3.0
170 stars 15 forks source link

Questions about the splits for argoverse2 #6

Closed haruishi43 closed 10 months ago

haruishi43 commented 10 months ago

Thanks for opensourcing your project.

In your argoverse2 preprocessing script argoverse_converter.py, I noticed that you removed 6 logs. It seems that there are only 4 logs that shouldn't be used for training models according to argoverse2 user guide. Are there any reasons for excluding additional logs?

Also, the text files seem to include 749/149 for train/val split which is different from the 700/150 split mentioned in the paper. Are the results obtained in the paper using the 749/149 split? Is the test set also being used for the additional logs for new train/val split?

yuantianyuan01 commented 10 months ago

Hi,

For the first question, the other two logs are removed due to NaN problem in map data. For the second question, we made a mistake in our paper. The new split on argoverse2 stands for the 749/149 split. Thank you for pointing it out and we will fix it in next revision.

haruishi43 commented 10 months ago

@yuantianyuan01 thanks for your response!

Following question, why did you removed the ~100 logs for argoverse2? In the appendix it mentions that you used all of the splits (~1000 logs) to generate train/val split. Were other samples removed due to NaN problems as well?

yuantianyuan01 commented 10 months ago

To make sure there's completely no overlap between training and validation set, we remove the logs on the borderline.