Add dataset table to the documentation

[calebrob6]

In the TorchGeo paper we have a table that lists the properties of some of the datasets we've added. This should be reproduced in the docs so that we have an overview of what all is available in the library.

I've just copied (more/less) the tables from the paper and haven't updated them with the datasets that have been implemented since.

Things to look into/questions:

• What other columns can we add for the Geospatial datasets? In the paper we've split it into Benchmark/Generic, however that's not really the organization that we have here. Perhaps we need a single table with a "Type" column?
• Adding hyperlinks to the citations for each dataset
• Updating the CSVs to be current

[adamjstewart]

What other columns can we add for the Geospatial datasets? In the paper we've split it into Benchmark/Generic, however that's not really the organization that we have here. Perhaps we need a single table with a "Type" column?

This starts to re-raise the question of how we should categorize datasets. We keep flip-flopping on this and it's leading to a lot of inconsistencies. Functionally, I think the biggest distinction is between geospatial datasets (have geospatial metadata) and non-geospatial datasets. In terms of attributes that we might want to list in this table, the biggest distinction is between benchmark datasets (bot input image and target labels) and non-benchmark datasets. The division in the docs doesn't necessarily need to match the base class division.

I'm leaning towards splitting the docs into benchmark vs. non-benchmark and keeping geospatial vs. non-geospatial as a base class distinction only. I've also been thinking about renaming VisionDataset to NonGeoDataset. Thoughts?

Adding hyperlinks to the citations for each dataset

Torchvision does this for their models and it always confuses me because I expect the hyperlink to take me to the model class definition, not the citation. I would prefer to have hyperlinks to class definitions and then the class definition contains a hyperlink to the citation. Thoughts?

Updating the CSVs to be current

We'll have to remind people to update this table every time they add a new dataset.

[calebrob6]

Dataset naming stuff

NonGeoDataset sounds great to me. And I'm fine with having a "benchmark dataset" table that doesn't align with how the classes are organized.

I would prefer to have hyperlinks to class definitions and then the class definition contains a hyperlink to the citation. Thoughts?

Fine with me -- rows in the table should definitely link somewhere.

We'll have to remind people to update this table every time they add a new dataset.

Yep! That's fine with me. I can also add instructions to the contributing page in this PR.

[ashnair1]

While we're on the topic, what about datasets like SpaceNet where they're pre-chipped to be like VisionDatasets but do contain geospatial metadata? In that case, it was the organisation of the dataset that informed the decision to make it a VisionDataset (query by integer index and not bounding box) not its lack of geospatial metadata.

Kind of lies in between geo-vs-vision

[adamjstewart]

@ashnair1 my current plan is to someday convert all of those to GeoDatasets (#83). The only thing holding us back at the moment is #409. I'm also planning on adding a new sampler (maybe PreChippedGeoSampler?) that doesn't require the user to specify the epoch length or patch size and instead gathers this directly from the dataset r-tree index. This will make them almost as simple as a VisionDataset but way more powerful since you can combine them with other datasets.

[adamjstewart]

In terms of filenames, there isn't a ton of consistency. We list "Geospatial Datasets" in "generic_datasets.csv" and "Non-geospatial Datasets" in "non_geo_datasets.csv". I'm honestly not sure what to call them anymore, and we've gone back and forth for a while. We should figure out how to make these more consistent in our docs/API. This doesn't necessarily need to happen in this PR, just pointing out the inconsistencies here.

[adamjstewart]

This looks great! Will take a closer look later. Since we first created the docs, torchvision's docs have completely changed. They now have a short page with just the dataset tables and then separate pages for each dataset. I actually kind of like that format, and it may allow us to skip the step of adding the dataset to datasets.rst. This doesn't change anything in this PR, but it may make this PR more important in the future.

[calebrob6]

Alright, ready for round 2