C-GQA dataset: Some pairs in train_pairs.txt don't have image examples in the training set

[vkhoi]

Thank you for the new dataset. I have one question about it. I find 133 pairs in train_pairs.txt that don't have any image examples in the training set. Some examples of them are: "angry person", "long_sleeved shirt", "open drawer", "small duck". Can you check this out?

[ferjad]

@vkhoi @HeimingX
Apologies for the late update on this. We found this issue and have corrected it in an updated version of the dataset. You can download the corrected c-gqa from this link.
The dataset stats and the new numbers for cge can be found in this arxiv preprint.
All the paper conclusions and ranking wrt to baseline methods still hold true. We have thoroughly checked for possible issues in this updated version but please let us know if you find any other problems.

[ToneLi]

are you sure this link can be opened? https://s3.mlcloud.uni-tuebingen.de/czsl/cgqa-updated.zip

[mancinimassimiliano]

Hello @ToneLi! Our servers are currently under maintenance (see #12). I contacted the admins and they mention they should be back online on Monday. I will let you know as soon as it is back online.

[mancinimassimiliano]

Update @ToneLi : the server is up again, so the dataset can be downloaded. Let us know if you encounter any issues.

[IshanManchanda]

Hi @mancinimassimiliano, I think the server is currently unreachable as I am unable to download the dataset. Can you confirm?

[mancinimassimiliano]

Hello @IshanManchanda! Yes, I confirm this. There was maintenance scheduled for today. I will check with the admins in case the problem is not fixed by tomorrow.

[IshanManchanda]

Hi @mancinimassimiliano, the file still seems to be down. Can you please check?

[mancinimassimiliano]

@IshanManchanda the dataset should be back online again (sorry, it took more time than expected).

[yzou2]

Thanks for providing the great dataset. What's the license of the c-gqa dataset?

[mancinimassimiliano]

Hello @yzou2! For the dataset license, @ferjad can provide a more precise answer.

[ferjad]

@yzou2 we follow the same license as the original GQA dataset https://creativecommons.org/licenses/by/4.0/