Dataset Size

[ctensmeyer]

In the ECCV 2020 paper, the size of PubTabNet is said to be 568K images + HTML. When I downloaded the v2 PubTabNet, I count 519030 image files and 509892 lines in the PubTabNet_2.0.0.jsonl file (presumably, the test split GT is omitted).

Can you explain this discrepancy and what effect it has on the ability of researchers to reproduce or fairly compare with your ECCV paper?

[ajjimeno]

Hi Christopher, As you mention there is a subset that is being used in the following challenge as test set: https://icdar2021.org/competitions/competition-on-scientific-literature-parsing/

…

On Wed, Feb 3, 2021 at 3:28 AM Christopher Tensmeyer < ***@***.***> wrote: In the ECCV 2020 paper, the size of PubTabNet is said to be 568K images + HTML. When I downloaded the v2 PubTabNet, I count 519030 image files and 509892 lines in the PubTabNet_2.0.0.jsonl file (presumably, the test split GT is omitted). Can you explain this discrepancy and what effect it has on the ability of researchers to reproduce or fairly compare with your ECCV paper? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#20>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AA6BZDPFJT4BRX6F5YEDZ7TS5AR4VANCNFSM4W67GCNQ> .

[ctensmeyer]

After the competition is completed, will the full 568K dataset be released to allow comparison with the ECCV paper?

[ajjimeno]

That is the intention. There is time till the 31st of March to participate in the challenge, in case you are interested.

…

On Fri, Feb 5, 2021 at 11:28 AM Christopher Tensmeyer < ***@***.***> wrote: After the competition is completed, will the full 568K dataset be released to allow comparison with the ECCV paper? — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#20 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AA6BZDIA4RTI6XMGTVPJ25TS5M3RJANCNFSM4W67GCNQ> .

[gjgjh]

Has the test split GT been released? Where can I find the download link? Thanks in advance.

[ajjimeno]

The test split should be available here: https://github.com/ajjimeno/icdar-task-b

…

On Tue, Jan 14, 2025 at 6:35 PM gjgjh ***@***.***> wrote: Has the test split GT been released? Where can I find the download link? Thanks in advance. — Reply to this email directly, view it on GitHub <#20 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AA6BZDI5HNPSTPJJ2HLDPQ32KXCOPAVCNFSM6AAAAABVGGZKMKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKOJRGUYTGNBWGM> . You are receiving this because you commented.Message ID: ***@***.***>

[gjgjh]

The test split should be available here: https://github.com/ajjimeno/icdar-task-b
…
On Tue, Jan 14, 2025 at 6:35 PM gjgjh @.> wrote: Has the test split GT been released? Where can I find the download link? Thanks in advance. — Reply to this email directly, view it on GitHub <#20 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AA6BZDI5HNPSTPJJ2HLDPQ32KXCOPAVCNFSM6AAAAABVGGZKMKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKOJRGUYTGNBWGM . You are receiving this because you commented.Message ID: @.>

Thank you for your reply. However, I feel that the images in the icdar-task-b dataset and the original PubTabNet test set do not match. For example, images like PMC509252_004_00.png and PMC514524_008_00.png from the PubTabNet test set cannot be found in the icdar-task-b dataset. Additionally, the icdar-task-b dataset contains 9064 samples, whereas the original PubTabNet test set has 9138 samples.
Nevertheless, the styles of the two datasets are quite similar and could be used as an alternative. I'm wondering if in academic work, people use icdar-task-b as the test set for PubTabNet?