search

RozDavid / LanguageGroundedSemseg

Implementation for ECCV 2022 paper Language-Grounded Indoor 3D Semantic Segmentation in the Wild
98 stars 14 forks source link

Acquiring the updated scannetv2-labels.combined.tsv #5

Closed LiyaoTang closed 1 year ago

LiyaoTang commented 1 year ago

Hi David,

Thanks for your data set release, it's amazing. However, when I tried to acquire the updated tsv file for ScanNet200, I simply re-run the download script with --label_map option and found the downloaded tsv file is the same as the one I already have.

For your convenience, the md5sum of the newly downloaded tsv file and the old file I have are the same, which is ea58dc6921c6a51569e3d3cf77fec182.

So, do I need to get the new download script to get the new tsv file? i.e., to fill another TOS form?

Best, Liyao Tang

RozDavid commented 1 year ago

Hey @LiyaoTang,

No need to fill anything, you have the correct file already based on the md5sum :) I dont know when you got it for the first time, but we updated the label map file on the server a couple of month earlier already. The other thing, there were only a few rows different in this and the older v2 label map, so maybe that's why you didn't notice the difference.

Kind regards, David

LiyaoTang commented 1 year ago

Hi @RozDavid,

Thanks for your prompt reply.

I’ve also considered the case that I have already got the new one. However, the thing that confuses me is that, it seems that several classes in the mapping have the same nyu40 id.

Or, do I need to use the id from another column, probably the "id" I guess?

Best, Liyao

RozDavid commented 1 year ago

Hi,

You are right, we use the id instead of _nyuid to avoid limitation to the predefined categories. You can also look into the preprocessing code here that might help undestanding the table.

Regards, David

LiyaoTang commented 1 year ago

Hi,

Thank you for the link, that's much clearer.

However, as I'm reading through the ScanNet200 instruction from the repo of ScanNet, I find this confusing.

Does it mean the benchmark would not include those labels (the VALID_CLASS_IDS_200_TRAIN_ONLY) for evaluation on test set?

Thanks again.

Best, Liyao

RozDavid commented 1 year ago

Hi Liyao,

So, this set of categories only corresponding to the differences between train and val sets. We didn't make it public which categories might or might not be present on the test set, so you are free to ignore that.

Regards, David

LiyaoTang commented 1 year ago

Ok, I see. Now everything is clear.

Thank you very much!

Cheers, Liyao