Added CUB2011 dataset

[vadimkantorov]

For testing some metric learning methods I created a wrapper for CUB2011 dataset (inheriting from CIFAR10 for download functionality and from ImageFolder for directory parsing).

The dataset is small, so it'd be possible to serve it from memory, but haven't needed it yet. Also not sure about the good practice on integrity hashes, just put there whole archive hash and a few image hashes.

Let me know what you think!

[szagoruyko]

can you add annotations?

[vadimkantorov]

I'll rename the class and will fix lint errors.

About the annotations, every image seems to have "15 Part Locations, 312 Binary Attributes, 1 Bounding Box".
Do we have an accepted format for bounding boxes? (x1-y1-x2-y2?)

I will check if an image always have 15 part locations, that would simplify the encoding.

[vadimkantorov]

VOC pull request indeed uses "[xmin, ymin, xmax, ymax, ind]" format so I'll stick to it.

[varunagrawal]

Is this PR still relevant? @vadimkantorov seems to have deleted his torchvision fork.

[vadimkantorov]

I am not working on this. Please go ahead if you wish to pick this up (my original CUB2011 dataset code is here: https://github.com/vadimkantorov/metriclearningbench/blob/master/cub2011.py )

Some things you may want to look at:

1. Refactor super-hacky class CUB2011(ImageFolder, CIFAR10) way to reuse dataset download steps. Maybe now there's a better API for that.
2. Actually implement "15 Part Locations, 312 Binary Attributes, 1 Bounding Box" loading (but maybe the CUB2011MetricLearning would suffice?)

[pmeier]

@vadimkantorov

I am not working on this.

Closing this in favor of #4126.