iWildCam potential dataset

[martinjaggi]

i think it would add a lot of value to have one dataset with more than a handful of domains/centers, as well as possibly one non-medical one

iWildCam would tick both boxes, with 323 centers

it's already part of wilds: https://wilds.stanford.edu/datasets/ so probably wouldn't be too much work to integrate (and licence already clear as well)

refs:
The iWildCam 2021 Competition Dataset
https://arxiv.org/abs/2105.03494

[jeandut]

Hello @martinjaggi this is indeed a suitable candidate. However I have a few problems with it:

• I would consider 323 as cross-device (more than 100), we could obviously group cameras by geo-information if it is available. Of course this is debatable.
• it is not healthcare related, which is good in general but considering the time we have and the other datasets we need to find at least a second dataset that is not healthcare so that it is coherent. This is again debatable maybe as you say one is sufficient.