Random sized patches support for other non-geospatial datamodules

[nilsleh]

Summary

In #851 it was suggested that datamodules with variable sized input images should implement a random size crop from tile to patches during training and have a keyword argument to specify the desired patch_size.

When looking through our other Non-geospatial datamodules, there are other cases that might be good to add the same support, just because the default images are very large and not suitable for batches even though all images have the same dimensions:

• GID-15, defaullt image size 6800x7200 (Add datamodule for GID-15 dataset #928)
• Inria Aerial Image Labeling, default image size 5000x5000
• DeepGlobeLandCoverDataModule, default image size is 2,448x2,448 (Add random crop logic to DeepGlobeLandCover Datamodule #876)
• Potsdam, default image size is 6000x6000 (Add crop logic to Potsdam2D datamodule #929)

Maybe, there can also be a threshold for future datamodule additions: for example if image size is greater than X, implement the random patch crop logic for that datamodule.

Rationale

I think it would lower the threshold of using these datasets. Since datamodules are supposed to make it very quick and easy for users to conduct segmentation experiments, the patch_size logic would make it a bit more easier to practically use the datamodules right away.

Implementation

Implement a logic like in the OSCDDatamodule.

Alternatives

No response

Additional information

No response

[ashnair1]

The InriaAerialImageLabelling datamodule already supports extraction of random crops.