Proposal for lazy-loaded annotations.

[LinasKo]

This branch contains a proposal for lazy-loading annotations for DetectionsDataset.

It's been tested scarcely - enough to show that memory usage does not grow when loading a large YOLO dataset. The main purpose is to have a code reference for further discussions.

Description

• A LazyDict data structure calls a loading function when values are retrieved
• This enables us to conceptualize a lazy_detections objects, which store masks as polygons until loaded.
• Such polygons are placed in data along with resolution_xy
• The dataset can take either lazy-loaded or normal Detections
• So far, only yolo_annotations_to_detections was changed to produce lazy values.
• The dataset can still function as before - the main different users will see is the datatype for annotations input: Union[Dict[str, Detections], LazyDict[str, Detections, Detections].

Type of change

Please delete options that are not relevant.

• [x] Bug fix (non-breaking change which fixes an issue)
• [x] New feature (non-breaking change which adds functionality)
• [x] This change requires a documentation update

How has this change been tested, please provide a testcase or example of how you tested the change?

pytest + loaded yolo dataset with it, comparing memory use and outputs with non-lazy method.

The memory use does not grow, and the results are the same.

I did not test this thoroughly so far.

Any specific deployment considerations

For example, documentation changes, usability, usage/costs, secrets, etc.

Docs

• [ ] Docs updated? What were the changes: