We have been asked by participants whether it is permitted to use additional, currently non-public labels in track 2 of the MIDOG challenge. We want to use this chance to clarify this topic: In track 2, the participants are allowed to use additional data, given that it is publicly available. Data that is fully automatically derived from existing data* is considered part of the method and thus permitted as well.
If you want to make use of additional manual labels that you or your organization has access to or labels generated by models that were trained on private data, this is only possible under the following condition:
- These additional labels are made available to the general public latest until July 19th (one month prior to the start of the final test phase of the challenge). Should you want to do this, we ask you to publish the data on zenodo with an accompanying github repository that gives an insight into how to use those labels
- You notify the organizers of the challenge briefly about it, so we can make this information available to everyone.
Should participants make use of private data/labels during training and not make it publicly available until this deadline, we will have to disqualify the approach.
Thanks for being part of our challenge and we are looking forward to a great workshop!
The Organizers
* If in doubt, ask yourself: Are all all steps & all models within this pipeline built/trained using only publicly available datasets? If the answer is yes, you are good to go. If you are still in doubt, ask the organizers.
Thanks Marc and team for the great job.
I suppose you comment on “Data that is fully automatically derived from existing data* is considered part of the method and thus permitted as well.” is only valid for the tack 2 of the challenge, right?
Does this mean that our approach for the last year challenge, which used NuClick generated masks, can not be used for the track 1 of the challenge? I would appreciate if you can clarify.
Best,
Mostafa
Dear Mostafa,
as NuClick requires the use of additional data (which it was trained upon), it is part of the track 2 of the challenge. Part 1 is only methods that require no additional data or models trained with additional data.
Best,
Marc