Stitching feature request: Transform non-geospatial model inference back into geospatial space.

[EricKeenan]

Summary

GridGeoSampler allows for efficient imagery chipping in preparation for ML-based inference. My problem is converting these non-geospatial inference results back into geospatial space. E.g. an inference file in geojson format that contains model inference bounding boxes, segmentations , etc.

Does this functionality already exist in TorchGeo? If not, how can we make this happen?

Rationale

From my perspective, this would enable machine learning experts to more easily interact with geospatial data.

Implementation

I have not thought deeply about implementation.

Alternatives

Not that I am aware of. Although there do seem to be related issues and PRs, e.g. https://github.com/microsoft/torchgeo/issues/1407.

Additional information

I would be interested in contributing to a solution. However, I'm not quite sure how to get started.

[adamjstewart]

Yes, #1407 is likely going to be the limiting factor here. I think the first step is to figure out how to keep things like CRS and transform in the sample batch. Maybe @adriantre has additional thoughts.