LandCover.ai: add GeoDataset version

[adrianboguszewski]

This PR renames LandCoverAI class to LandCoverAINonGeo (benchmark dataset, created with split.py script) and introduces the abstract base LandCoverAI class and LandCoverAIGeo (raster dataset).

This is my proposition, but I am open to the discussion of how the structure, inheritance, and naming should look.

[adrianboguszewski]

@microsoft-github-policy-service agree

[adamjstewart]

This looks correct modulo some documentation hiccups. Let me mull over the class names. Some options:

• LandCoverAINonGeo and LandCoverAIGeo
• LandCoverAI and LandCoverAIGeo
• LandCoverAIBenchmark and LandCoverAIUncurated

I think this will be the first dataset where we have both a geo and non-geo version, so we have an opportunity to set a new standard naming scheme.

[calebrob6]

For naming, I prefer something like LandCoverAI for the GeoDataset and LandCoverAIFixed / LandCoverAIPatched for the NonGeoDataset.

[adamjstewart]

I think @isaaccorley said he prefers LandCoverAI for non-geo and LandCoverAIGeo for geo. Not sure how to come to agreement on the names. Would be great if we can decide on an approach that is consistent for all datasets, there are many others that could be both geo and non-geo.

[calebrob6]

Yeah on second thought I like that (LandCoverAI for non-geo and LandCoverAIGeo for geo) a bit better:

• Doesn't break existing code
• It makes the non-geo dataset seem like the default (which is likely what people want if they're just trying to compare to published numbers)

[adamjstewart]

The base class can be called LandCoverAIBase to disambiguate

[adrianboguszewski]

Thank you guys for the review and a valuable name discussion. backwards-incompatible label can be removed.