Add xView2 Dataset

[isaaccorley]

This PR adds the xView2 train/test set used in the xView2 challenge and described in "xBD: A Dataset for Assessing Building Damage from Satellite Imagery", Gupta et al. (2019)

This allows for a user to load the train/test 1024x1024 imagery and masks as a Vision Dataset, however in the homepage there are additional download links for non chipped imagery and the geojson labels which we should consider adding later as a Geo Dataset. This dataset cannot be automatically downloaded as you must register at the xView2 homepage and the download links expire after a few minutes.

• Added a draw_semantic_segmentation_mask method to torchgeo.datasets.utils which takes an image tensor, semantic mask tensor (h, w), alpha blend factor, and colormap and overlays the mask onto the image, returning a numpy array which can be used for plotting

TODO:
- plot examples

• add unit tests

Examples:

[isaaccorley]

Taken from https://pytorch.org/vision/master/generated/torchvision.utils.draw_segmentation_masks.html#torchvision.utils.draw_segmentation_masks

[isaaccorley]

@adamjstewart sphinx docs builds are failing due to the below errors. Any idea?

/home/runner/work/torchgeo/torchgeo/torchgeo/datasets/__init__.py:docstring of torchgeo.datasets.xview.XView2.plot:: WARNING: py:class reference target not found: matplotlib.figure.Figure
/home/runner/work/torchgeo/torchgeo/torchgeo/datasets/__init__.py:docstring of torchgeo.datasets.VisionClassificationDataset:1: WARNING: py:class reference target not found: torchvision.datasets.folder.ImageFolder

[calebrob6]

@isaaccorley this is from the intersphinx mapping, fix is here

torchgeo/docs/conf.py

Line 102 in 41502e1

"matplotlib": ("https://matplotlib.org/stable/", None),

(should be merged soon)

[calebrob6]

@adamjstewart @isaaccorley I think this is finished now.

Some notes:

• It seems like in many datasets we aren't handing the "checksum" argument properly (i.e. it is only checked in the code path that downloads the dataset).
• This dataset has pre- and post-disaster imagery with pre- and post-disaster building masks/damage labels, so it will not fit nicely with the pre-canned SemanticSegmentationTask.
• We should make it clear that this is the "Challenge training set" and "Challenge test set" data from the xView2 website. I'll do this with any last changes that need to happen.

[adamjstewart]

Suggested change

	@pytest.fixture(params=["train", "test"])
	@pytest.fixture(scope="class", params=["train", "test"])

[calebrob6]

This breaks the tests

[calebrob6]

ScopeMismatch: You tried to access the 'function' scoped fixture 'monkeypatch' with a 'class' scoped request object, involved factories