[feat] Addition of popular image retrieval benchmark datasets

[ir2718]

Hi,

this is a PR for issue #722. I've implemented four benchmark datasets: CUB-200, Cars196, INaturalist2018, and StanfordOnlineProducts. When using any of these datasets they will be downloaded directly and saved to the root directory, similar to the PyTorch dataset handling. Each of the implemented datasets inherits torch.utils.data.Dataset and can be used with dataloaders seamlessly. I've also added docs for each of the datasets implement, and a short overview of what users need to implemented if they want to add their own custom dataset. Tests for each of the datasets are also added. I've deliberately left out the __init__.py in tests/datasets, as each of the files has to be downloaded, and these can be pretty big (up to 130Gb).

@KevinMusgrave when you have time, please take a look and tell me if something requires changing.

[ir2718]

Are there any plans on merging this or would you like me to make some changes?

[KevinMusgrave]

Looks good I think, but it would be nice to get more "official" links for SOP and Cars, but especially for SOP because I don't like having people download an anonymous-looking google drive link 😅

[ir2718]

Regarding the Cars dataset, unfortunately the original website is down. Even torchvision mentions this in their docs.

For SOP the used link is actually the official link. You can see the same link here in the readme of their repository.

[KevinMusgrave]

Thanks @ir2718!