Proposed updates to OSCD dataset

[robmarkcole]

Summary

• Proposal 1: Following EuroSATDataModule initialised with 3 bands gives error: mean length and number of channels do not match #1634 the OSCD dataset should be updated to support bands=["B04", "B03", "B02"] rather than the current bands in ["rgb", "all"]

• Proposal 2: OSCD is a change dataset and currently the before/after images are concatenated. However as discussed here it is preferable to return these separately. Required updates to __getitem__ below:

        image1 = self._load_image(files["images1"])
        image2 = self._load_image(files["images2"])
        mask = self._load_target(str(files["mask"]))

        if self.transforms is not None:
            image = torch.cat([image1, image2])
            sample = {"image": image, "mask": mask}
            sample = self.transforms(sample)
            idx = sample["image"].shape[0] // 2
            image1 = sample["image"][:idx]
            image2 = sample["image"][idx:]

        sample = {"image1": image1, "image2": image2, "mask": mask}
        return sample

Rationale

Harmonise the use of bands arg and prepare for a ChangeDetectionTask

Implementation

No response

Alternatives

No response

Additional information

No response

[adamjstewart]

Agree with both of these changes. I think in #289 we decided to standardize on bands being a list of strings of band names, so any datasets that don't follow this are outdated.

[robmarkcole]

Possibly we should also consider mirroring this dataset to huggingface as the native download is sloooow

[adamjstewart]

You'll need to figure out what license the data is distributed under and ensure that it allows for redistribution.