Add VHR10DataModule

[ashnair1]

• Added plot function for VHR10
• Added test

Since this is a CocoDataset it might be better to subclass from torchvision's CocoDataset instead of NonGeoDataset. Thoughts?

Plot sample (split="positive"):

Had to bump minimum version of pyococotools to 2.0.1 to support draw_bbox arg.

[adamjstewart]

This looks great. A few suggestions to make it even better:

1. Can we also plot the bounding box of each object? VHR10 was originally an OD dataset but someone later added IS masks as well.
2. Plotting masks and boxes for each object should be optional, we'll likely have OD and IS trainers long before we have an OD+IS trainer
3. Can we also plot the label of each object? Want to make sure we aren't detecting the wrong object
4. Can we plot each instance of the same class in the same color? This helps visualize the former point better
5. Can we also plot the confidence interval of each object?
6. Can the alpha value of the mask be passed as input to the plot method?

A sample plot from VHR10 with what I'm looking for (prediction, not ground truth):

Probably better to have the label/confidence above the box instead of inside, but close enough.

This was created using the following notebook, mostly using code from torchvision:
geog527train.zip

Fun fact: VHR10 was actually the first dataset I added to TorchGeo.

[ashnair1]

1. Can we also plot the bounding box of each object? VHR10 was originally an OD dataset but someone later added IS masks as well.
2. Plotting masks and boxes for each object should be optional, we'll likely have OD and IS trainers long before we have an OD+IS trainer
3. Can we also plot the label of each object? Want to make sure we aren't detecting the wrong object
4. Can we plot each instance of the same class in the same color? This helps visualize the former point better
5. Can we also plot the confidence interval of each object?
6. Can the alpha value of the mask be passed as input to the plot method?

Regarding point 1, by passing draw_boxes=True in the plot function, boxes will be drawn along with the mask.

[ashnair1]

That being said, I get where you're coming from. Plots from pycocotools are not as visually impressive, customizable or as informative as the second plot.

I've written out a new plot function that does this. Still ironing out some details but here is a sample. (Recommend opening the image in another window for better clarity).

Left (Masks and Boxes)
Center (Only Masks with mask alpha = 0.2)
Right (Only boxes)

[ashnair1]

Training a object detector on VHR10 now and plots look encouraging. Might need bump up the figure size.

[calebrob6]

@ashnair1 maybe try throwing in a plt.tightlayout() plt.tight_layout() before returning plt.gcf()

[ashnair1]

This PR will be split into smaller PRs once I make sure everything works as expected.

[adamjstewart]

Note that the datamodule will need to be refactored once #992 is merged. But that should hopefully make it significantly simpler.

[adamjstewart]

Closed by #847 and #1082